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

Update gpu_hist algorithm (#2901)

Browse Source
This commit is contained in:
Rory Mitchell
2017-11-27 13:44:24 +13:00
committed by GitHub
parent 24f527a1c0
commit c55f14668e
7 changed files with 771 additions and 1751 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
src/learner.cc
View File

@@ -111,7 +111,6 @@ struct LearnerTrainParam : public dmlc::Parameter<LearnerTrainParam> {
.add_enum("hist", 3)
.add_enum("gpu_exact", 4)
.add_enum("gpu_hist", 5)
.add_enum("gpu_hist_experimental", 6)
.describe("Choice of tree construction method.");
DMLC_DECLARE_FIELD(test_flag).set_default("").describe(
"Internal test flag");
@@ -188,14 +187,6 @@ class LearnerImpl : public Learner {
if (cfg_.count("predictor") == 0) {
cfg_["predictor"] = "gpu_predictor";
}
} else if (tparam.tree_method == 6) {
this->AssertGPUSupport();
if (cfg_.count("updater") == 0) {
cfg_["updater"] = "grow_gpu_hist_experimental,prune";
}
if (cfg_.count("predictor") == 0) {
cfg_["predictor"] = "gpu_predictor";
}
}
}
@@ -468,7 +459,7 @@ class LearnerImpl : public Learner {
// if not, initialize the column access.
inline void LazyInitDMatrix(DMatrix* p_train) {
if (tparam.tree_method == 3 || tparam.tree_method == 4 ||
tparam.tree_method == 5 || tparam.tree_method == 6) {
tparam.tree_method == 5) {
return;
}

1
src/tree/tree_updater.cc
View File

@@ -35,7 +35,6 @@ DMLC_REGISTRY_LINK_TAG(updater_sync);
#ifdef XGBOOST_USE_CUDA
DMLC_REGISTRY_LINK_TAG(updater_gpu);
DMLC_REGISTRY_LINK_TAG(updater_gpu_hist);
DMLC_REGISTRY_LINK_TAG(updater_gpu_hist_experimental);
#endif
} // namespace tree
} // namespace xgboost

1589
src/tree/updater_gpu_hist.cu
View File

File diff suppressed because it is too large Load Diff

899
src/tree/updater_gpu_hist_experimental.cu
View File

@@ -1,899 +0,0 @@
/*!
* Copyright 2017 XGBoost contributors
*/
#include <thrust/execution_policy.h>
#include <thrust/reduce.h>
#include <thrust/sequence.h>
#include <xgboost/tree_updater.h>
#include <algorithm>
#include <memory>
#include <queue>
#include <utility>
#include <vector>
#include "../common/compressed_iterator.h"
#include "../common/device_helpers.cuh"
#include "../common/hist_util.h"
#include "../common/timer.h"
#include "param.h"
#include "updater_gpu_common.cuh"
namespace xgboost {
namespace tree {
DMLC_REGISTRY_FILE_TAG(updater_gpu_hist_experimental);
typedef bst_gpair_precise gpair_sum_t;
template <int BLOCK_THREADS, typename reduce_t, typename temp_storage_t>
__device__ gpair_sum_t ReduceFeature(const gpair_sum_t* begin,
const gpair_sum_t* end,
temp_storage_t* temp_storage) {
__shared__ cub::Uninitialized<gpair_sum_t> uninitialized_sum;
gpair_sum_t& shared_sum = uninitialized_sum.Alias();
gpair_sum_t local_sum = gpair_sum_t();
for (auto itr = begin; itr < end; itr += BLOCK_THREADS) {
bool thread_active = itr + threadIdx.x < end;
// Scan histogram
gpair_sum_t bin = thread_active ? *(itr + threadIdx.x) : gpair_sum_t();
local_sum += reduce_t(temp_storage->sum_reduce).Reduce(bin, cub::Sum());
}
if (threadIdx.x == 0) {
shared_sum = local_sum;
}
__syncthreads();
return shared_sum;
}
template <int BLOCK_THREADS, typename reduce_t, typename scan_t,
typename max_reduce_t, typename temp_storage_t>
__device__ void EvaluateFeature(int fidx, const gpair_sum_t* hist,
const int* feature_segments, float min_fvalue,
const float* gidx_fvalue_map,
DeviceSplitCandidate* best_split,
const DeviceNodeStats& node,
const GPUTrainingParam& param,
temp_storage_t* temp_storage) {
int gidx_begin = feature_segments[fidx];
int gidx_end = feature_segments[fidx + 1];
gpair_sum_t feature_sum = ReduceFeature<BLOCK_THREADS, reduce_t>(
hist + gidx_begin, hist + gidx_end, temp_storage);
auto prefix_op = SumCallbackOp<gpair_sum_t>();
for (int scan_begin = gidx_begin; scan_begin < gidx_end;
scan_begin += BLOCK_THREADS) {
bool thread_active = scan_begin + threadIdx.x < gidx_end;
gpair_sum_t bin =
thread_active ? hist[scan_begin + threadIdx.x] : gpair_sum_t();
scan_t(temp_storage->scan).ExclusiveScan(bin, bin, cub::Sum(), prefix_op);
// Calculate gain
gpair_sum_t parent_sum = gpair_sum_t(node.sum_gradients);
gpair_sum_t missing = parent_sum - feature_sum;
bool missing_left = true;
const float null_gain = -FLT_MAX;
float gain = null_gain;
if (thread_active) {
gain = loss_chg_missing(bin, missing, parent_sum, node.root_gain, param,
missing_left);
}
__syncthreads();
// Find thread with best gain
cub::KeyValuePair<int, float> tuple(threadIdx.x, gain);
cub::KeyValuePair<int, float> best =
max_reduce_t(temp_storage->max_reduce).Reduce(tuple, cub::ArgMax());
__shared__ cub::KeyValuePair<int, float> block_max;
if (threadIdx.x == 0) {
block_max = best;
}
__syncthreads();
// Best thread updates split
if (threadIdx.x == block_max.key) {
int gidx = scan_begin + threadIdx.x;
float fvalue =
gidx == gidx_begin ? min_fvalue : gidx_fvalue_map[gidx - 1];
gpair_sum_t left = missing_left ? bin + missing : bin;
gpair_sum_t right = parent_sum - left;
best_split->Update(gain, missing_left ? LeftDir : RightDir, fvalue, fidx,
left, right, param);
}
__syncthreads();
}
}
template <int BLOCK_THREADS>
__global__ void evaluate_split_kernel(
const gpair_sum_t* d_hist, int nidx, uint64_t n_features,
DeviceNodeStats nodes, const int* d_feature_segments,
const float* d_fidx_min_map, const float* d_gidx_fvalue_map,
GPUTrainingParam gpu_param, DeviceSplitCandidate* d_split) {
typedef cub::KeyValuePair<int, float> ArgMaxT;
typedef cub::BlockScan<gpair_sum_t, BLOCK_THREADS, cub::BLOCK_SCAN_WARP_SCANS>
BlockScanT;
typedef cub::BlockReduce<ArgMaxT, BLOCK_THREADS> MaxReduceT;
typedef cub::BlockReduce<gpair_sum_t, BLOCK_THREADS> SumReduceT;
union TempStorage {
typename BlockScanT::TempStorage scan;
typename MaxReduceT::TempStorage max_reduce;
typename SumReduceT::TempStorage sum_reduce;
};
__shared__ cub::Uninitialized<DeviceSplitCandidate> uninitialized_split;
DeviceSplitCandidate& best_split = uninitialized_split.Alias();
__shared__ TempStorage temp_storage;
if (threadIdx.x == 0) {
best_split = DeviceSplitCandidate();
}
__syncthreads();
auto fidx = blockIdx.x;
EvaluateFeature<BLOCK_THREADS, SumReduceT, BlockScanT, MaxReduceT>(
fidx, d_hist, d_feature_segments, d_fidx_min_map[fidx], d_gidx_fvalue_map,
&best_split, nodes, gpu_param, &temp_storage);
__syncthreads();
if (threadIdx.x == 0) {
// Record best loss
d_split[fidx] = best_split;
}
}
// Find a gidx value for a given feature otherwise return -1 if not found
template <typename gidx_iter_t>
__device__ int BinarySearchRow(bst_uint begin, bst_uint end, gidx_iter_t data,
int fidx_begin, int fidx_end) {
bst_uint previous_middle = UINT32_MAX;
while (end != begin) {
auto middle = begin + (end - begin) / 2;
if (middle == previous_middle) {
break;
}
previous_middle = middle;
auto gidx = data[middle];
if (gidx >= fidx_begin && gidx < fidx_end) {
return gidx;
} else if (gidx < fidx_begin) {
begin = middle;
} else {
end = middle;
}
}
// Value is missing
return -1;
}
struct DeviceHistogram {
dh::bulk_allocator<dh::memory_type::DEVICE> ba;
dh::dvec<gpair_sum_t> data;
int n_bins;
void Init(int device_idx, int max_nodes, int n_bins, bool silent) {
this->n_bins = n_bins;
ba.allocate(device_idx, silent, &data, size_t(max_nodes) * size_t(n_bins));
}
void Reset() { data.fill(gpair_sum_t()); }
gpair_sum_t* GetHistPtr(int nidx) { return data.data() + nidx * n_bins; }
void PrintNidx(int nidx) const {
auto h_data = data.as_vector();
std::cout << "nidx " << nidx << ":\n";
for (int i = n_bins * nidx; i < n_bins * (nidx + 1); i++) {
std::cout << h_data[i] << " ";
}
std::cout << "\n";
}
};
// Manage memory for a single GPU
struct DeviceShard {
struct Segment {
size_t begin;
size_t end;
Segment() : begin(0), end(0) {}
Segment(size_t begin, size_t end) : begin(begin), end(end) {
CHECK_GE(end, begin);
}
size_t Size() const { return end - begin; }
};
int device_idx;
int normalised_device_idx; // Device index counting from param.gpu_id
dh::bulk_allocator<dh::memory_type::DEVICE> ba;
dh::dvec<common::compressed_byte_t> gidx_buffer;
dh::dvec<bst_gpair> gpair;
dh::dvec2<bst_uint> ridx; // Row index relative to this shard
dh::dvec2<int> position;
std::vector<Segment> ridx_segments;
dh::dvec<int> feature_segments;
dh::dvec<float> gidx_fvalue_map;
dh::dvec<float> min_fvalue;
std::vector<bst_gpair> node_sum_gradients;
common::CompressedIterator<uint32_t> gidx;
int row_stride;
bst_uint row_begin_idx; // The row offset for this shard
bst_uint row_end_idx;
bst_uint n_rows;
int n_bins;
int null_gidx_value;
DeviceHistogram hist;
TrainParam param;
int64_t* tmp_pinned; // Small amount of staging memory
std::vector<cudaStream_t> streams;
dh::CubMemory temp_memory;
DeviceShard(int device_idx, int normalised_device_idx,
const common::GHistIndexMatrix& gmat, bst_uint row_begin,
bst_uint row_end, int n_bins, TrainParam param)
: device_idx(device_idx),
normalised_device_idx(normalised_device_idx),
row_begin_idx(row_begin),
row_end_idx(row_end),
n_rows(row_end - row_begin),
n_bins(n_bins),
null_gidx_value(n_bins),
param(param) {
// Convert to ELLPACK matrix representation
int max_elements_row = 0;
for (auto i = row_begin; i < row_end; i++) {
max_elements_row =
(std::max)(max_elements_row,
static_cast<int>(gmat.row_ptr[i + 1] - gmat.row_ptr[i]));
}
row_stride = max_elements_row;
std::vector<int> ellpack_matrix(row_stride * n_rows, null_gidx_value);
for (auto i = row_begin; i < row_end; i++) {
int row_count = 0;
for (auto j = gmat.row_ptr[i]; j < gmat.row_ptr[i + 1]; j++) {
ellpack_matrix[(i - row_begin) * row_stride + row_count] =
gmat.index[j];
row_count++;
}
}
// Allocate
int num_symbols = n_bins + 1;
size_t compressed_size_bytes =
common::CompressedBufferWriter::CalculateBufferSize(
ellpack_matrix.size(), num_symbols);
int max_nodes =
param.max_leaves > 0 ? param.max_leaves * 2 : n_nodes(param.max_depth);
ba.allocate(device_idx, param.silent, &gidx_buffer, compressed_size_bytes,
&gpair, n_rows, &ridx, n_rows, &position, n_rows,
&feature_segments, gmat.cut->row_ptr.size(), &gidx_fvalue_map,
gmat.cut->cut.size(), &min_fvalue, gmat.cut->min_val.size());
gidx_fvalue_map = gmat.cut->cut;
min_fvalue = gmat.cut->min_val;
feature_segments = gmat.cut->row_ptr;
node_sum_gradients.resize(max_nodes);
ridx_segments.resize(max_nodes);
// Compress gidx
common::CompressedBufferWriter cbw(num_symbols);
std::vector<common::compressed_byte_t> host_buffer(gidx_buffer.size());
cbw.Write(host_buffer.data(), ellpack_matrix.begin(), ellpack_matrix.end());
gidx_buffer = host_buffer;
gidx =
common::CompressedIterator<uint32_t>(gidx_buffer.data(), num_symbols);
common::CompressedIterator<uint32_t> ci_host(host_buffer.data(),
num_symbols);
// Init histogram
hist.Init(device_idx, max_nodes, gmat.cut->row_ptr.back(), param.silent);
dh::safe_cuda(cudaMallocHost(&tmp_pinned, sizeof(int64_t)));
}
~DeviceShard() {
for (auto& stream : streams) {
dh::safe_cuda(cudaStreamDestroy(stream));
}
dh::safe_cuda(cudaFreeHost(tmp_pinned));
}
// Get vector of at least n initialised streams
std::vector<cudaStream_t>& GetStreams(int n) {
if (n > streams.size()) {
for (auto& stream : streams) {
dh::safe_cuda(cudaStreamDestroy(stream));
}
streams.clear();
streams.resize(n);
for (auto& stream : streams) {
dh::safe_cuda(cudaStreamCreate(&stream));
}
}
return streams;
}
// Reset values for each update iteration
void Reset(const std::vector<bst_gpair>& host_gpair) {
dh::safe_cuda(cudaSetDevice(device_idx));
position.current_dvec().fill(0);
std::fill(node_sum_gradients.begin(), node_sum_gradients.end(),
bst_gpair());
thrust::sequence(ridx.current_dvec().tbegin(), ridx.current_dvec().tend());
std::fill(ridx_segments.begin(), ridx_segments.end(), Segment(0, 0));
ridx_segments.front() = Segment(0, ridx.size());
this->gpair.copy(host_gpair.begin() + row_begin_idx,
host_gpair.begin() + row_end_idx);
subsample_gpair(&gpair, param.subsample, row_begin_idx);
hist.Reset();
}
void BuildHist(int nidx) {
auto segment = ridx_segments[nidx];
auto d_node_hist = hist.GetHistPtr(nidx);
auto d_gidx = gidx;
auto d_ridx = ridx.current();
auto d_gpair = gpair.data();
auto row_stride = this->row_stride;
auto null_gidx_value = this->null_gidx_value;
auto n_elements = segment.Size() * row_stride;
dh::launch_n(device_idx, n_elements, [=] __device__(size_t idx) {
int ridx = d_ridx[(idx / row_stride) + segment.begin];
int gidx = d_gidx[ridx * row_stride + idx % row_stride];
if (gidx != null_gidx_value) {
AtomicAddGpair(d_node_hist + gidx, d_gpair[ridx]);
}
});
}
void SubtractionTrick(int nidx_parent, int nidx_histogram,
int nidx_subtraction) {
auto d_node_hist_parent = hist.GetHistPtr(nidx_parent);
auto d_node_hist_histogram = hist.GetHistPtr(nidx_histogram);
auto d_node_hist_subtraction = hist.GetHistPtr(nidx_subtraction);
dh::launch_n(device_idx, hist.n_bins, [=] __device__(size_t idx) {
d_node_hist_subtraction[idx] =
d_node_hist_parent[idx] - d_node_hist_histogram[idx];
});
}
__device__ void CountLeft(int64_t* d_count, int val, int left_nidx) {
unsigned ballot = __ballot(val == left_nidx);
if (threadIdx.x % 32 == 0) {
atomicAdd(reinterpret_cast<unsigned long long*>(d_count), // NOLINT
static_cast<unsigned long long>(__popc(ballot))); // NOLINT
}
}
void UpdatePosition(int nidx, int left_nidx, int right_nidx, int fidx,
int split_gidx, bool default_dir_left, bool is_dense,
int fidx_begin, int fidx_end) {
dh::safe_cuda(cudaSetDevice(device_idx));
temp_memory.LazyAllocate(sizeof(int64_t));
auto d_left_count = temp_memory.Pointer<int64_t>();
dh::safe_cuda(cudaMemset(d_left_count, 0, sizeof(int64_t)));
auto segment = ridx_segments[nidx];
auto d_ridx = ridx.current();
auto d_position = position.current();
auto d_gidx = gidx;
auto row_stride = this->row_stride;
dh::launch_n<1, 512>(
device_idx, segment.Size(), [=] __device__(bst_uint idx) {
idx += segment.begin;
auto ridx = d_ridx[idx];
auto row_begin = row_stride * ridx;
auto row_end = row_begin + row_stride;
auto gidx = -1;
if (is_dense) {
gidx = d_gidx[row_begin + fidx];
} else {
gidx = BinarySearchRow(row_begin, row_end, d_gidx, fidx_begin,
fidx_end);
}
int position;
if (gidx >= 0) {
// Feature is found
position = gidx <= split_gidx ? left_nidx : right_nidx;
} else {
// Feature is missing
position = default_dir_left ? left_nidx : right_nidx;
}
CountLeft(d_left_count, position, left_nidx);
d_position[idx] = position;
});
dh::safe_cuda(cudaMemcpy(tmp_pinned, d_left_count, sizeof(int64_t),
cudaMemcpyDeviceToHost));
auto left_count = *tmp_pinned;
SortPosition(segment, left_nidx, right_nidx);
// dh::safe_cuda(cudaStreamSynchronize(stream));
ridx_segments[left_nidx] =
Segment(segment.begin, segment.begin + left_count);
ridx_segments[right_nidx] =
Segment(segment.begin + left_count, segment.end);
}
void SortPosition(const Segment& segment, int left_nidx, int right_nidx) {
int min_bits = 0;
int max_bits = static_cast<int>(
std::ceil(std::log2((std::max)(left_nidx, right_nidx) + 1)));
size_t temp_storage_bytes = 0;
cub::DeviceRadixSort::SortPairs(
nullptr, temp_storage_bytes, position.current() + segment.begin,
position.other() + segment.begin, ridx.current() + segment.begin,
ridx.other() + segment.begin, segment.Size(), min_bits, max_bits);
temp_memory.LazyAllocate(temp_storage_bytes);
cub::DeviceRadixSort::SortPairs(
temp_memory.d_temp_storage, temp_memory.temp_storage_bytes,
position.current() + segment.begin, position.other() + segment.begin,
ridx.current() + segment.begin, ridx.other() + segment.begin,
segment.Size(), min_bits, max_bits);
dh::safe_cuda(cudaMemcpy(
position.current() + segment.begin, position.other() + segment.begin,
segment.Size() * sizeof(int), cudaMemcpyDeviceToDevice));
dh::safe_cuda(cudaMemcpy(
ridx.current() + segment.begin, ridx.other() + segment.begin,
segment.Size() * sizeof(bst_uint), cudaMemcpyDeviceToDevice));
}
};
class GPUHistMakerExperimental : public TreeUpdater {
public:
struct ExpandEntry;
GPUHistMakerExperimental() : initialised(false) {}
~GPUHistMakerExperimental() {}
void Init(
const std::vector<std::pair<std::string, std::string>>& args) override {
param.InitAllowUnknown(args);
CHECK(param.n_gpus != 0) << "Must have at least one device";
n_devices = param.n_gpus;
dh::check_compute_capability();
if (param.grow_policy == TrainParam::kLossGuide) {
qexpand_.reset(new ExpandQueue(loss_guide));
} else {
qexpand_.reset(new ExpandQueue(depth_wise));
}
monitor.Init("updater_gpu_hist_experimental", param.debug_verbose);
}
void Update(const std::vector<bst_gpair>& gpair, DMatrix* dmat,
const std::vector<RegTree*>& trees) override {
monitor.Start("Update");
GradStats::CheckInfo(dmat->info());
// rescale learning rate according to size of trees
float lr = param.learning_rate;
param.learning_rate = lr / trees.size();
// build tree
try {
for (size_t i = 0; i < trees.size(); ++i) {
this->UpdateTree(gpair, dmat, trees[i]);
}
} catch (const std::exception& e) {
LOG(FATAL) << "GPU plugin exception: " << e.what() << std::endl;
}
param.learning_rate = lr;
monitor.Stop("Update");
}
void InitDataOnce(DMatrix* dmat) {
info = &dmat->info();
monitor.Start("Quantiles");
hmat_.Init(dmat, param.max_bin);
gmat_.cut = &hmat_;
gmat_.Init(dmat);
monitor.Stop("Quantiles");
n_bins = hmat_.row_ptr.back();
int n_devices = dh::n_devices(param.n_gpus, info->num_row);
bst_uint row_begin = 0;
bst_uint shard_size =
std::ceil(static_cast<double>(info->num_row) / n_devices);
std::vector<int> dList(n_devices);
for (int d_idx = 0; d_idx < n_devices; ++d_idx) {
int device_idx = (param.gpu_id + d_idx) % dh::n_visible_devices();
dList[d_idx] = device_idx;
}
reducer.Init(dList);
// Partition input matrix into row segments
std::vector<size_t> row_segments;
shards.resize(n_devices);
row_segments.push_back(0);
for (int d_idx = 0; d_idx < n_devices; ++d_idx) {
bst_uint row_end =
std::min(static_cast<size_t>(row_begin + shard_size), info->num_row);
row_segments.push_back(row_end);
row_begin = row_end;
}
// Create device shards
omp_set_num_threads(shards.size());
#pragma omp parallel
{
auto cpu_thread_id = omp_get_thread_num();
shards[cpu_thread_id] = std::unique_ptr<DeviceShard>(
new DeviceShard(dList[cpu_thread_id], cpu_thread_id, gmat_,
row_segments[cpu_thread_id],
row_segments[cpu_thread_id + 1], n_bins, param));
}
initialised = true;
}
void InitData(const std::vector<bst_gpair>& gpair, DMatrix* dmat,
const RegTree& tree) {
monitor.Start("InitDataOnce");
if (!initialised) {
this->InitDataOnce(dmat);
}
monitor.Stop("InitDataOnce");
column_sampler.Init(info->num_col, param);
// Copy gpair & reset memory
monitor.Start("InitDataReset");
omp_set_num_threads(shards.size());
#pragma omp parallel
{
auto cpu_thread_id = omp_get_thread_num();
shards[cpu_thread_id]->Reset(gpair);
}
monitor.Stop("InitDataReset");
}
void AllReduceHist(int nidx) {
for (auto& shard : shards) {
auto d_node_hist = shard->hist.GetHistPtr(nidx);
reducer.AllReduceSum(
shard->normalised_device_idx,
reinterpret_cast<gpair_sum_t::value_t*>(d_node_hist),
reinterpret_cast<gpair_sum_t::value_t*>(d_node_hist),
n_bins * (sizeof(gpair_sum_t) / sizeof(gpair_sum_t::value_t)));
}
reducer.Synchronize();
}
void BuildHistLeftRight(int nidx_parent, int nidx_left, int nidx_right) {
size_t left_node_max_elements = 0;
size_t right_node_max_elements = 0;
for (auto& shard : shards) {
left_node_max_elements = (std::max)(
left_node_max_elements, shard->ridx_segments[nidx_left].Size());
right_node_max_elements = (std::max)(
right_node_max_elements, shard->ridx_segments[nidx_right].Size());
}
auto build_hist_nidx = nidx_left;
auto subtraction_trick_nidx = nidx_right;
if (right_node_max_elements < left_node_max_elements) {
build_hist_nidx = nidx_right;
subtraction_trick_nidx = nidx_left;
}
for (auto& shard : shards) {
shard->BuildHist(build_hist_nidx);
}
this->AllReduceHist(build_hist_nidx);
for (auto& shard : shards) {
shard->SubtractionTrick(nidx_parent, build_hist_nidx,
subtraction_trick_nidx);
}
}
// Returns best loss
std::vector<DeviceSplitCandidate> EvaluateSplits(
const std::vector<int>& nidx_set, RegTree* p_tree) {
auto columns = info->num_col;
std::vector<DeviceSplitCandidate> best_splits(nidx_set.size());
std::vector<DeviceSplitCandidate> candidate_splits(nidx_set.size() *
columns);
// Use first device
auto& shard = shards.front();
dh::safe_cuda(cudaSetDevice(shard->device_idx));
shard->temp_memory.LazyAllocate(sizeof(DeviceSplitCandidate) * columns *
nidx_set.size());
auto d_split = shard->temp_memory.Pointer<DeviceSplitCandidate>();
auto& streams = shard->GetStreams(static_cast<int>(nidx_set.size()));
// Use streams to process nodes concurrently
for (auto i = 0; i < nidx_set.size(); i++) {
auto nidx = nidx_set[i];
DeviceNodeStats node(shard->node_sum_gradients[nidx], nidx, param);
const int BLOCK_THREADS = 256;
evaluate_split_kernel<BLOCK_THREADS>
<<<uint32_t(columns), BLOCK_THREADS, 0, streams[i]>>>(
shard->hist.GetHistPtr(nidx), nidx, info->num_col, node,
shard->feature_segments.data(), shard->min_fvalue.data(),
shard->gidx_fvalue_map.data(), GPUTrainingParam(param),
d_split + i * columns);
}
dh::safe_cuda(
cudaMemcpy(candidate_splits.data(), shard->temp_memory.d_temp_storage,
sizeof(DeviceSplitCandidate) * columns * nidx_set.size(),
cudaMemcpyDeviceToHost));
for (auto i = 0; i < nidx_set.size(); i++) {
auto nidx = nidx_set[i];
DeviceSplitCandidate nidx_best;
for (auto fidx = 0; fidx < columns; fidx++) {
auto& candidate = candidate_splits[i * columns + fidx];
if (column_sampler.ColumnUsed(candidate.findex,
p_tree->GetDepth(nidx))) {
nidx_best.Update(candidate_splits[i * columns + fidx], param);
}
}
best_splits[i] = nidx_best;
}
return std::move(best_splits);
}
void InitRoot(const std::vector<bst_gpair>& gpair, RegTree* p_tree) {
auto root_nidx = 0;
// Sum gradients
std::vector<bst_gpair> tmp_sums(shards.size());
omp_set_num_threads(shards.size());
#pragma omp parallel
{
auto cpu_thread_id = omp_get_thread_num();
dh::safe_cuda(cudaSetDevice(shards[cpu_thread_id]->device_idx));
tmp_sums[cpu_thread_id] =
thrust::reduce(thrust::cuda::par(shards[cpu_thread_id]->temp_memory),
shards[cpu_thread_id]->gpair.tbegin(),
shards[cpu_thread_id]->gpair.tend());
}
auto sum_gradient =
std::accumulate(tmp_sums.begin(), tmp_sums.end(), bst_gpair());
// Generate root histogram
for (auto& shard : shards) {
shard->BuildHist(root_nidx);
}
this->AllReduceHist(root_nidx);
// Remember root stats
p_tree->stat(root_nidx).sum_hess = sum_gradient.GetHess();
p_tree->stat(root_nidx).base_weight = CalcWeight(param, sum_gradient);
// Store sum gradients
for (auto& shard : shards) {
shard->node_sum_gradients[root_nidx] = sum_gradient;
}
// Generate first split
auto splits = this->EvaluateSplits({root_nidx}, p_tree);
qexpand_->push(
ExpandEntry(root_nidx, p_tree->GetDepth(root_nidx), splits.front(), 0));
}
void UpdatePosition(const ExpandEntry& candidate, RegTree* p_tree) {
auto nidx = candidate.nid;
auto left_nidx = (*p_tree)[nidx].cleft();
auto right_nidx = (*p_tree)[nidx].cright();
// convert floating-point split_pt into corresponding bin_id
// split_cond = -1 indicates that split_pt is less than all known cut points
auto split_gidx = -1;
auto fidx = candidate.split.findex;
auto default_dir_left = candidate.split.dir == LeftDir;
auto fidx_begin = hmat_.row_ptr[fidx];
auto fidx_end = hmat_.row_ptr[fidx + 1];
for (auto i = fidx_begin; i < fidx_end; ++i) {
if (candidate.split.fvalue == hmat_.cut[i]) {
split_gidx = static_cast<int32_t>(i);
}
}
auto is_dense = info->num_nonzero == info->num_row * info->num_col;
omp_set_num_threads(shards.size());
#pragma omp parallel
{
auto cpu_thread_id = omp_get_thread_num();
shards[cpu_thread_id]->UpdatePosition(nidx, left_nidx, right_nidx, fidx,
split_gidx, default_dir_left,
is_dense, fidx_begin, fidx_end);
}
}
void ApplySplit(const ExpandEntry& candidate, RegTree* p_tree) {
// Add new leaves
RegTree& tree = *p_tree;
tree.AddChilds(candidate.nid);
auto& parent = tree[candidate.nid];
parent.set_split(candidate.split.findex, candidate.split.fvalue,
candidate.split.dir == LeftDir);
tree.stat(candidate.nid).loss_chg = candidate.split.loss_chg;
// Configure left child
auto left_weight = CalcWeight(param, candidate.split.left_sum);
tree[parent.cleft()].set_leaf(left_weight * param.learning_rate, 0);
tree.stat(parent.cleft()).base_weight = left_weight;
tree.stat(parent.cleft()).sum_hess = candidate.split.left_sum.GetHess();
// Configure right child
auto right_weight = CalcWeight(param, candidate.split.right_sum);
tree[parent.cright()].set_leaf(right_weight * param.learning_rate, 0);
tree.stat(parent.cright()).base_weight = right_weight;
tree.stat(parent.cright()).sum_hess = candidate.split.right_sum.GetHess();
// Store sum gradients
for (auto& shard : shards) {
shard->node_sum_gradients[parent.cleft()] = candidate.split.left_sum;
shard->node_sum_gradients[parent.cright()] = candidate.split.right_sum;
}
this->UpdatePosition(candidate, p_tree);
}
void UpdateTree(const std::vector<bst_gpair>& gpair, DMatrix* p_fmat,
RegTree* p_tree) {
// Temporarily store number of threads so we can change it back later
int nthread = omp_get_max_threads();
auto& tree = *p_tree;
monitor.Start("InitData");
this->InitData(gpair, p_fmat, *p_tree);
monitor.Stop("InitData");
monitor.Start("InitRoot");
this->InitRoot(gpair, p_tree);
monitor.Stop("InitRoot");
auto timestamp = qexpand_->size();
auto num_leaves = 1;
while (!qexpand_->empty()) {
auto candidate = qexpand_->top();
qexpand_->pop();
if (!candidate.IsValid(param, num_leaves)) continue;
// std::cout << candidate;
monitor.Start("ApplySplit");
this->ApplySplit(candidate, p_tree);
monitor.Stop("ApplySplit");
num_leaves++;
auto left_child_nidx = tree[candidate.nid].cleft();
auto right_child_nidx = tree[candidate.nid].cright();
// Only create child entries if needed
if (ExpandEntry::ChildIsValid(param, tree.GetDepth(left_child_nidx),
num_leaves)) {
monitor.Start("BuildHist");
this->BuildHistLeftRight(candidate.nid, left_child_nidx,
right_child_nidx);
monitor.Stop("BuildHist");
monitor.Start("EvaluateSplits");
auto splits =
this->EvaluateSplits({left_child_nidx, right_child_nidx}, p_tree);
qexpand_->push(ExpandEntry(left_child_nidx,
tree.GetDepth(left_child_nidx), splits[0],
timestamp++));
qexpand_->push(ExpandEntry(right_child_nidx,
tree.GetDepth(right_child_nidx), splits[1],
timestamp++));
monitor.Stop("EvaluateSplits");
}
}
// Reset omp num threads
omp_set_num_threads(nthread);
}
struct ExpandEntry {
int nid;
int depth;
DeviceSplitCandidate split;
uint64_t timestamp;
ExpandEntry(int nid, int depth, const DeviceSplitCandidate& split,
uint64_t timestamp)
: nid(nid), depth(depth), split(split), timestamp(timestamp) {}
bool IsValid(const TrainParam& param, int num_leaves) const {
if (split.loss_chg <= rt_eps) return false;
if (param.max_depth > 0 && depth == param.max_depth) return false;
if (param.max_leaves > 0 && num_leaves == param.max_leaves) return false;
return true;
}
static bool ChildIsValid(const TrainParam& param, int depth,
int num_leaves) {
if (param.max_depth > 0 && depth == param.max_depth) return false;
if (param.max_leaves > 0 && num_leaves == param.max_leaves) return false;
return true;
}
friend std::ostream& operator<<(std::ostream& os, const ExpandEntry& e) {
os << "ExpandEntry: \n";
os << "nidx: " << e.nid << "\n";
os << "depth: " << e.depth << "\n";
os << "loss: " << e.split.loss_chg << "\n";
os << "left_sum: " << e.split.left_sum << "\n";
os << "right_sum: " << e.split.right_sum << "\n";
return os;
}
};
inline static bool depth_wise(ExpandEntry lhs, ExpandEntry rhs) {
if (lhs.depth == rhs.depth) {
return lhs.timestamp > rhs.timestamp; // favor small timestamp
} else {
return lhs.depth > rhs.depth; // favor small depth
}
}
inline static bool loss_guide(ExpandEntry lhs, ExpandEntry rhs) {
if (lhs.split.loss_chg == rhs.split.loss_chg) {
return lhs.timestamp > rhs.timestamp; // favor small timestamp
} else {
return lhs.split.loss_chg < rhs.split.loss_chg; // favor large loss_chg
}
}
TrainParam param;
common::HistCutMatrix hmat_;
common::GHistIndexMatrix gmat_;
MetaInfo* info;
bool initialised;
int n_devices;
int n_bins;
std::vector<std::unique_ptr<DeviceShard>> shards;
ColumnSampler column_sampler;
typedef std::priority_queue<ExpandEntry, std::vector<ExpandEntry>,
std::function<bool(ExpandEntry, ExpandEntry)>>
ExpandQueue;
std::unique_ptr<ExpandQueue> qexpand_;
common::Monitor monitor;
dh::AllReducer reducer;
};
XGBOOST_REGISTER_TREE_UPDATER(GPUHistMakerExperimental,
"grow_gpu_hist_experimental")
.describe("Grow tree with GPU.")
.set_body([]() { return new GPUHistMakerExperimental(); });
} // namespace tree
} // namespace xgboost