[EM] Support mmap backed ellpack. (#10602)

- Support resource view in ellpack.
- Define the CUDA version of MMAP resource.
- Define the CUDA version of malloc resource.
- Refactor cuda runtime API wrappers, and add memory access related wrappers.
- gather windows macros into a single header.
This commit is contained in:
Jiaming Yuan
2024-07-18 08:20:21 +08:00
committed by GitHub
parent e9fbce9791
commit 292bb677e5
59 changed files with 889 additions and 646 deletions

View File

@@ -1,5 +1,5 @@
/**
* Copyright 2020-2023, XGBoost Contributors
* Copyright 2020-2024, XGBoost Contributors
*/
#include <gtest/gtest.h>
@@ -102,19 +102,17 @@ TEST(GradientBasedSampler, NoSamplingExternalMemory) {
EXPECT_EQ(sample.gpair.data(), gpair.DevicePointer());
EXPECT_EQ(sampled_page->n_rows, kRows);
std::vector<common::CompressedByteT> buffer(sampled_page->gidx_buffer.HostVector());
common::CompressedIterator<common::CompressedByteT>
ci(buffer.data(), sampled_page->NumSymbols());
std::vector<common::CompressedByteT> h_gidx_buffer;
auto h_accessor = sampled_page->GetHostAccessor(&ctx, &h_gidx_buffer);
size_t offset = 0;
std::size_t offset = 0;
for (auto& batch : dmat->GetBatches<EllpackPage>(&ctx, param)) {
auto page = batch.Impl();
std::vector<common::CompressedByteT> page_buffer(page->gidx_buffer.HostVector());
common::CompressedIterator<common::CompressedByteT>
page_ci(page_buffer.data(), page->NumSymbols());
std::vector<common::CompressedByteT> h_page_gidx_buffer;
auto page_accessor = page->GetHostAccessor(&ctx, &h_page_gidx_buffer);
size_t num_elements = page->n_rows * page->row_stride;
for (size_t i = 0; i < num_elements; i++) {
EXPECT_EQ(ci[i + offset], page_ci[i]);
EXPECT_EQ(h_accessor.gidx_iter[i + offset], page_accessor.gidx_iter[i]);
}
offset += num_elements;
}

View File

@@ -328,8 +328,7 @@ class HistogramExternalMemoryTest : public ::testing::TestWithParam<std::tuple<f
for (auto const& page : p_fmat->GetBatches<SparsePage>()) {
concat.Push(page);
}
EllpackPageImpl page{
ctx.Device(), cuts, concat, p_fmat->IsDense(), p_fmat->Info().num_col_, {}};
EllpackPageImpl page{&ctx, cuts, concat, p_fmat->IsDense(), p_fmat->Info().num_col_, {}};
auto ridx = partitioner.GetRows(0);
auto d_histogram = dh::ToSpan(single_hist);
DeviceHistogramBuilder builder;

View File

@@ -81,6 +81,7 @@ std::vector<GradientPairPrecise> GetHostHistGpair() {
template <typename GradientSumT>
void TestBuildHist(bool use_shared_memory_histograms) {
int const kNRows = 16, kNCols = 8;
Context ctx{MakeCUDACtx(0)};
TrainParam param;
Args args{
@@ -89,9 +90,8 @@ void TestBuildHist(bool use_shared_memory_histograms) {
};
param.Init(args);
auto page = BuildEllpackPage(kNRows, kNCols);
auto page = BuildEllpackPage(&ctx, kNRows, kNCols);
BatchParam batch_param{};
Context ctx{MakeCUDACtx(0)};
auto cs = std::make_shared<common::ColumnSampler>(0);
GPUHistMakerDevice maker(&ctx, /*is_external_memory=*/false, {}, kNRows, param, cs, kNCols,
batch_param, MetaInfo());
@@ -105,7 +105,6 @@ void TestBuildHist(bool use_shared_memory_histograms) {
}
gpair.SetDevice(ctx.Device());
thrust::host_vector<common::CompressedByteT> h_gidx_buffer(page->gidx_buffer.HostVector());
maker.row_partitioner = std::make_unique<RowPartitioner>(&ctx, kNRows, 0);
maker.hist.Init(ctx.Device(), page->Cuts().TotalBins());
@@ -198,14 +197,12 @@ void TestHistogramIndexImpl() {
auto grad = GenerateRandomGradients(kNRows);
grad.SetDevice(DeviceOrd::CUDA(0));
maker->Reset(&grad, hist_maker_dmat.get(), kNCols);
std::vector<common::CompressedByteT> h_gidx_buffer(maker->page->gidx_buffer.HostVector());
const auto &maker_ext = hist_maker_ext.maker;
maker_ext->Reset(&grad, hist_maker_ext_dmat.get(), kNCols);
std::vector<common::CompressedByteT> h_gidx_buffer_ext(maker_ext->page->gidx_buffer.HostVector());
ASSERT_EQ(maker->page->Cuts().TotalBins(), maker_ext->page->Cuts().TotalBins());
ASSERT_EQ(maker->page->gidx_buffer.Size(), maker_ext->page->gidx_buffer.Size());
ASSERT_EQ(maker->page->gidx_buffer.size(), maker_ext->page->gidx_buffer.size());
}
TEST(GpuHist, TestHistogramIndex) {