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

46
tests/cpp/data/test_ellpack_page.cu
View File

@@ -19,23 +19,19 @@ TEST(EllpackPage, EmptyDMatrix) {
auto dmat = *CreateDMatrix(kNRows, kNCols, kSparsity);
auto& page = *dmat->GetBatches<EllpackPage>({0, kMaxBin, kGpuBatchNRows}).begin();
auto impl = page.Impl();
ASSERT_EQ(impl->matrix.info.feature_segments.size(), 1);
ASSERT_EQ(impl->matrix.info.min_fvalue.size(), 0);
ASSERT_EQ(impl->matrix.info.gidx_fvalue_map.size(), 0);
ASSERT_EQ(impl->matrix.info.row_stride, 0);
ASSERT_EQ(impl->matrix.info.n_bins, 0);
ASSERT_EQ(impl->gidx_buffer.size(), 4);
ASSERT_EQ(impl->row_stride, 0);
ASSERT_EQ(impl->cuts_.TotalBins(), 0);
ASSERT_EQ(impl->gidx_buffer.Size(), 4);
}
TEST(EllpackPage, BuildGidxDense) {
int constexpr kNRows = 16, kNCols = 8;
auto page = BuildEllpackPage(kNRows, kNCols);
std::vector<common::CompressedByteT> h_gidx_buffer(page->gidx_buffer.size());
dh::CopyDeviceSpanToVector(&h_gidx_buffer, page->gidx_buffer);
common::CompressedIterator<uint32_t> gidx(h_gidx_buffer.data(), 25);
std::vector<common::CompressedByteT> h_gidx_buffer(page->gidx_buffer.HostVector());
common::CompressedIterator<uint32_t> gidx(h_gidx_buffer.data(), page->NumSymbols());
ASSERT_EQ(page->matrix.info.row_stride, kNCols);
ASSERT_EQ(page->row_stride, kNCols);
std::vector<uint32_t> solution = {
0, 3, 8, 9, 14, 17, 20, 21,
@@ -64,11 +60,10 @@ TEST(EllpackPage, BuildGidxSparse) {
int constexpr kNRows = 16, kNCols = 8;
auto page = BuildEllpackPage(kNRows, kNCols, 0.9f);
std::vector<common::CompressedByteT> h_gidx_buffer(page->gidx_buffer.size());
dh::CopyDeviceSpanToVector(&h_gidx_buffer, page->gidx_buffer);
std::vector<common::CompressedByteT> h_gidx_buffer(page->gidx_buffer.HostVector());
common::CompressedIterator<uint32_t> gidx(h_gidx_buffer.data(), 25);
ASSERT_LE(page->matrix.info.row_stride, 3);
ASSERT_LE(page->row_stride, 3);
// row_stride = 3, 16 rows, 48 entries for ELLPack
std::vector<uint32_t> solution = {
@@ -76,16 +71,16 @@ TEST(EllpackPage, BuildGidxSparse) {
24, 24, 24, 24, 24, 5, 24, 24, 0, 16, 24, 15, 24, 24, 24, 24,
24, 7, 14, 16, 4, 24, 24, 24, 24, 24, 9, 24, 24, 1, 24, 24
};
for (size_t i = 0; i < kNRows * page->matrix.info.row_stride; ++i) {
for (size_t i = 0; i < kNRows * page->row_stride; ++i) {
ASSERT_EQ(solution[i], gidx[i]);
}
}
struct ReadRowFunction {
EllpackMatrix matrix;
EllpackDeviceAccessor matrix;
int row;
bst_float* row_data_d;
ReadRowFunction(EllpackMatrix matrix, int row, bst_float* row_data_d)
ReadRowFunction(EllpackDeviceAccessor matrix, int row, bst_float* row_data_d)
: matrix(std::move(matrix)), row(row), row_data_d(row_data_d) {}
__device__ void operator()(size_t col) {
@@ -110,7 +105,8 @@ TEST(EllpackPage, Copy) {
auto page = (*dmat->GetBatches<EllpackPage>(param).begin()).Impl();
// Create an empty result page.
EllpackPageImpl result(0, page->matrix.info, kRows);
EllpackPageImpl result(0, page->cuts_, page->is_dense, page->row_stride,
kRows);
// Copy batch pages into the result page.
size_t offset = 0;
@@ -126,13 +122,13 @@ TEST(EllpackPage, Copy) {
std::vector<bst_float> row_result(kCols);
for (auto& page : dmat->GetBatches<EllpackPage>(param)) {
auto impl = page.Impl();
EXPECT_EQ(impl->matrix.base_rowid, current_row);
EXPECT_EQ(impl->base_rowid, current_row);
for (size_t i = 0; i < impl->Size(); i++) {
dh::LaunchN(0, kCols, ReadRowFunction(impl->matrix, current_row, row_d.data().get()));
dh::LaunchN(0, kCols, ReadRowFunction(impl->GetDeviceAccessor(0), current_row, row_d.data().get()));
thrust::copy(row_d.begin(), row_d.end(), row.begin());
dh::LaunchN(0, kCols, ReadRowFunction(result.matrix, current_row, row_result_d.data().get()));
dh::LaunchN(0, kCols, ReadRowFunction(result.GetDeviceAccessor(0), current_row, row_result_d.data().get()));
thrust::copy(row_result_d.begin(), row_result_d.end(), row_result.begin());
EXPECT_EQ(row, row_result);
@@ -155,7 +151,8 @@ TEST(EllpackPage, Compact) {
auto page = (*dmat->GetBatches<EllpackPage>(param).begin()).Impl();
// Create an empty result page.
EllpackPageImpl result(0, page->matrix.info, kCompactedRows);
EllpackPageImpl result(0, page->cuts_, page->is_dense, page->row_stride,
kCompactedRows);
// Compact batch pages into the result page.
std::vector<size_t> row_indexes_h {
@@ -174,7 +171,7 @@ TEST(EllpackPage, Compact) {
std::vector<bst_float> row_result(kCols);
for (auto& page : dmat->GetBatches<EllpackPage>(param)) {
auto impl = page.Impl();
EXPECT_EQ(impl->matrix.base_rowid, current_row);
EXPECT_EQ(impl->base_rowid, current_row);
for (size_t i = 0; i < impl->Size(); i++) {
size_t compacted_row = row_indexes_h[current_row];
@@ -183,11 +180,12 @@ TEST(EllpackPage, Compact) {
continue;
}
dh::LaunchN(0, kCols, ReadRowFunction(impl->matrix, current_row, row_d.data().get()));
dh::LaunchN(0, kCols, ReadRowFunction(impl->GetDeviceAccessor(0), current_row, row_d.data().get()));
dh::safe_cuda (cudaDeviceSynchronize());
thrust::copy(row_d.begin(), row_d.end(), row.begin());
dh::LaunchN(0, kCols,
ReadRowFunction(result.matrix, compacted_row, row_result_d.data().get()));
ReadRowFunction(result.GetDeviceAccessor(0), compacted_row, row_result_d.data().get()));
thrust::copy(row_result_d.begin(), row_result_d.end(), row_result.begin());
EXPECT_EQ(row, row_result);

45
tests/cpp/data/test_sparse_page_dmatrix.cu
View File

@@ -3,6 +3,7 @@
#include <dmlc/filesystem.h>
#include "../helpers.h"
#include "../../../src/common/compressed_iterator.h"
#include "../../../src/data/ellpack_page.cuh"
namespace xgboost {
@@ -58,31 +59,29 @@ TEST(SparsePageDMatrix, EllpackPageContent) {
BatchParam param{0, 2, 0, 0};
auto impl = (*dmat->GetBatches<EllpackPage>(param).begin()).Impl();
EXPECT_EQ(impl->matrix.base_rowid, 0);
EXPECT_EQ(impl->matrix.n_rows, kRows);
EXPECT_FALSE(impl->matrix.info.is_dense);
EXPECT_EQ(impl->matrix.info.row_stride, 2);
EXPECT_EQ(impl->matrix.info.n_bins, 4);
EXPECT_EQ(impl->base_rowid, 0);
EXPECT_EQ(impl->n_rows, kRows);
EXPECT_FALSE(impl->is_dense);
EXPECT_EQ(impl->row_stride, 2);
EXPECT_EQ(impl->cuts_.TotalBins(), 4);
auto impl_ext = (*dmat_ext->GetBatches<EllpackPage>(param).begin()).Impl();
EXPECT_EQ(impl_ext->matrix.base_rowid, 0);
EXPECT_EQ(impl_ext->matrix.n_rows, kRows);
EXPECT_FALSE(impl_ext->matrix.info.is_dense);
EXPECT_EQ(impl_ext->matrix.info.row_stride, 2);
EXPECT_EQ(impl_ext->matrix.info.n_bins, 4);
EXPECT_EQ(impl_ext->base_rowid, 0);
EXPECT_EQ(impl_ext->n_rows, kRows);
EXPECT_FALSE(impl_ext->is_dense);
EXPECT_EQ(impl_ext->row_stride, 2);
EXPECT_EQ(impl_ext->cuts_.TotalBins(), 4);
std::vector<common::CompressedByteT> buffer(impl->gidx_buffer.size());
std::vector<common::CompressedByteT> buffer_ext(impl_ext->gidx_buffer.size());
dh::CopyDeviceSpanToVector(&buffer, impl->gidx_buffer);
dh::CopyDeviceSpanToVector(&buffer_ext, impl_ext->gidx_buffer);
std::vector<common::CompressedByteT> buffer(impl->gidx_buffer.HostVector());
std::vector<common::CompressedByteT> buffer_ext(impl_ext->gidx_buffer.HostVector());
EXPECT_EQ(buffer, buffer_ext);
}
struct ReadRowFunction {
EllpackMatrix matrix;
EllpackDeviceAccessor matrix;
int row;
bst_float* row_data_d;
ReadRowFunction(EllpackMatrix matrix, int row, bst_float* row_data_d)
ReadRowFunction(EllpackDeviceAccessor matrix, int row, bst_float* row_data_d)
: matrix(std::move(matrix)), row(row), row_data_d(row_data_d) {}
__device__ void operator()(size_t col) {
@@ -110,8 +109,8 @@ TEST(SparsePageDMatrix, MultipleEllpackPageContent) {
BatchParam param{0, kMaxBins, 0, kPageSize};
auto impl = (*dmat->GetBatches<EllpackPage>(param).begin()).Impl();
EXPECT_EQ(impl->matrix.base_rowid, 0);
EXPECT_EQ(impl->matrix.n_rows, kRows);
EXPECT_EQ(impl->base_rowid, 0);
EXPECT_EQ(impl->n_rows, kRows);
size_t current_row = 0;
thrust::device_vector<bst_float> row_d(kCols);
@@ -120,13 +119,13 @@ TEST(SparsePageDMatrix, MultipleEllpackPageContent) {
std::vector<bst_float> row_ext(kCols);
for (auto& page : dmat_ext->GetBatches<EllpackPage>(param)) {
auto impl_ext = page.Impl();
EXPECT_EQ(impl_ext->matrix.base_rowid, current_row);
EXPECT_EQ(impl_ext->base_rowid, current_row);
for (size_t i = 0; i < impl_ext->Size(); i++) {
dh::LaunchN(0, kCols, ReadRowFunction(impl->matrix, current_row, row_d.data().get()));
dh::LaunchN(0, kCols, ReadRowFunction(impl->GetDeviceAccessor(0), current_row, row_d.data().get()));
thrust::copy(row_d.begin(), row_d.end(), row.begin());
dh::LaunchN(0, kCols, ReadRowFunction(impl_ext->matrix, current_row, row_ext_d.data().get()));
dh::LaunchN(0, kCols, ReadRowFunction(impl_ext->GetDeviceAccessor(0), current_row, row_ext_d.data().get()));
thrust::copy(row_ext_d.begin(), row_ext_d.end(), row_ext.begin());
EXPECT_EQ(row, row_ext);
@@ -155,8 +154,8 @@ TEST(SparsePageDMatrix, EllpackPageMultipleLoops) {
size_t current_row = 0;
for (auto& page : dmat_ext->GetBatches<EllpackPage>(param)) {
auto impl_ext = page.Impl();
EXPECT_EQ(impl_ext->matrix.base_rowid, current_row);
current_row += impl_ext->matrix.n_rows;
EXPECT_EQ(impl_ext->base_rowid, current_row);
current_row += impl_ext->n_rows;
}
}