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

Partial rewrite EllpackPage (#5352)

Browse Source
This commit is contained in:
Rory Mitchell
2020-03-11 10:15:53 +13:00
committed by GitHub
parent 7a99f8f27f
commit 3ad4333b0e
23 changed files with 496 additions and 733 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
src/tree/gpu_hist/gradient_based_sampler.cu
View File

@@ -153,7 +153,8 @@ ExternalMemoryNoSampling::ExternalMemoryNoSampling(EllpackPageImpl* page,
size_t n_rows,
const BatchParam& batch_param)
: batch_param_(batch_param),
page_(new EllpackPageImpl(batch_param.gpu_id, page->matrix.info, n_rows)) {}
page_(new EllpackPageImpl(batch_param.gpu_id, page->cuts_, page->is_dense,
page->row_stride, n_rows)) {}
GradientBasedSample ExternalMemoryNoSampling::Sample(common::Span<GradientPair> gpair,
DMatrix* dmat) {
@@ -217,9 +218,9 @@ GradientBasedSample ExternalMemoryUniformSampling::Sample(common::Span<GradientP
// Create a new ELLPACK page with empty rows.
page_.reset(); // Release the device memory first before reallocating
page_.reset(new EllpackPageImpl(batch_param_.gpu_id,
original_page_->matrix.info,
sample_rows));
page_.reset(new EllpackPageImpl(
batch_param_.gpu_id, original_page_->cuts_, original_page_->is_dense,
original_page_->row_stride, sample_rows));
// Compact the ELLPACK pages into the single sample page.
thrust::fill(dh::tbegin(page_->gidx_buffer), dh::tend(page_->gidx_buffer), 0);
@@ -298,9 +299,9 @@ GradientBasedSample ExternalMemoryGradientBasedSampling::Sample(common::Span<Gra
// Create a new ELLPACK page with empty rows.
page_.reset(); // Release the device memory first before reallocating
page_.reset(new EllpackPageImpl(batch_param_.gpu_id,
original_page_->matrix.info,
sample_rows));
page_.reset(new EllpackPageImpl(batch_param_.gpu_id, original_page_->cuts_,
original_page_->is_dense,
original_page_->row_stride, sample_rows));
// Compact the ELLPACK pages into the single sample page.
thrust::fill(dh::tbegin(page_->gidx_buffer), dh::tend(page_->gidx_buffer), 0);
@@ -319,7 +320,7 @@ GradientBasedSampler::GradientBasedSampler(EllpackPageImpl* page,
monitor_.Init("gradient_based_sampler");
bool is_sampling = subsample < 1.0;
bool is_external_memory = page->matrix.n_rows != n_rows;
bool is_external_memory = page->n_rows != n_rows;
if (is_sampling) {
switch (sampling_method) {

22
src/tree/gpu_hist/histogram.cu
View File

@@ -101,7 +101,7 @@ template GradientPairPrecise CreateRoundingFactor(common::Span<GradientPair cons
template GradientPair CreateRoundingFactor(common::Span<GradientPair const> gpair);
template <typename GradientSumT>
__global__ void SharedMemHistKernel(xgboost::EllpackMatrix matrix,
__global__ void SharedMemHistKernel(EllpackDeviceAccessor matrix,
common::Span<const RowPartitioner::RowIndexT> d_ridx,
GradientSumT* __restrict__ d_node_hist,
const GradientPair* __restrict__ d_gpair,
@@ -112,14 +112,14 @@ __global__ void SharedMemHistKernel(xgboost::EllpackMatrix matrix,
extern __shared__ char smem[];
GradientSumT* smem_arr = reinterpret_cast<GradientSumT*>(smem); // NOLINT
if (use_shared_memory_histograms) {
dh::BlockFill(smem_arr, matrix.info.n_bins, GradientSumT());
dh::BlockFill(smem_arr, matrix.NumBins(), GradientSumT());
__syncthreads();
}
for (auto idx : dh::GridStrideRange(static_cast<size_t>(0), n_elements)) {
int ridx = d_ridx[idx / matrix.info.row_stride];
int ridx = d_ridx[idx / matrix.row_stride];
int gidx =
matrix.gidx_iter[ridx * matrix.info.row_stride + idx % matrix.info.row_stride];
if (gidx != matrix.info.n_bins) {
matrix.gidx_iter[ridx * matrix.row_stride + idx % matrix.row_stride];
if (gidx != matrix.NumBins()) {
GradientSumT truncated {
TruncateWithRoundingFactor<T>(rounding.GetGrad(), d_gpair[ridx].GetGrad()),
TruncateWithRoundingFactor<T>(rounding.GetHess(), d_gpair[ridx].GetHess()),
@@ -135,7 +135,7 @@ __global__ void SharedMemHistKernel(xgboost::EllpackMatrix matrix,
if (use_shared_memory_histograms) {
// Write shared memory back to global memory
__syncthreads();
for (auto i : dh::BlockStrideRange(static_cast<size_t>(0), matrix.info.n_bins)) {
for (auto i : dh::BlockStrideRange(static_cast<size_t>(0), matrix.NumBins())) {
GradientSumT truncated {
TruncateWithRoundingFactor<T>(rounding.GetGrad(), smem_arr[i].GetGrad()),
TruncateWithRoundingFactor<T>(rounding.GetHess(), smem_arr[i].GetHess()),
@@ -146,16 +146,16 @@ __global__ void SharedMemHistKernel(xgboost::EllpackMatrix matrix,
}
template <typename GradientSumT>
void BuildGradientHistogram(EllpackMatrix const& matrix,
void BuildGradientHistogram(EllpackDeviceAccessor const& matrix,
common::Span<GradientPair const> gpair,
common::Span<const uint32_t> d_ridx,
common::Span<GradientSumT> histogram,
GradientSumT rounding, bool shared) {
const size_t smem_size =
shared
? sizeof(GradientSumT) * matrix.info.n_bins
? sizeof(GradientSumT) * matrix.NumBins()
: 0;
auto n_elements = d_ridx.size() * matrix.info.row_stride;
auto n_elements = d_ridx.size() * matrix.row_stride;
uint32_t items_per_thread = 8;
uint32_t block_threads = 256;
@@ -168,14 +168,14 @@ void BuildGradientHistogram(EllpackMatrix const& matrix,
}
template void BuildGradientHistogram<GradientPair>(
EllpackMatrix const& matrix,
EllpackDeviceAccessor const& matrix,
common::Span<GradientPair const> gpair,
common::Span<const uint32_t> ridx,
common::Span<GradientPair> histogram,
GradientPair rounding, bool shared);
template void BuildGradientHistogram<GradientPairPrecise>(
EllpackMatrix const& matrix,
EllpackDeviceAccessor const& matrix,
common::Span<GradientPair const> gpair,
common::Span<const uint32_t> ridx,
common::Span<GradientPairPrecise> histogram,

2
src/tree/gpu_hist/histogram.cuh
View File

@@ -18,7 +18,7 @@ DEV_INLINE T TruncateWithRoundingFactor(T const rounding_factor, float const x)
}
template <typename GradientSumT>
void BuildGradientHistogram(EllpackMatrix const& matrix,
void BuildGradientHistogram(EllpackDeviceAccessor const& matrix,
common::Span<GradientPair const> gpair,
common::Span<const uint32_t> ridx,
common::Span<GradientSumT> histogram,