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

Fuse gpu_hist all-reduce calls where possible (#7867)

Browse Source
This commit is contained in:
Rory Mitchell
2022-05-17 13:27:50 +02:00
committed by GitHub
parent b41cf92dc2
commit 71d3b2e036
9 changed files with 234 additions and 185 deletions
Download Patch File Download Diff File Expand all files Collapse all files

45
tests/cpp/tree/gpu_hist/test_driver.cu
View File

@@ -6,41 +6,58 @@ namespace xgboost {
namespace tree {
TEST(GpuHist, DriverDepthWise) {
Driver<GPUExpandEntry> driver(TrainParam::kDepthWise);
TrainParam p;
p.InitAllowUnknown(Args{});
p.grow_policy = TrainParam::kDepthWise;
Driver<GPUExpandEntry> driver(p, 2);
EXPECT_TRUE(driver.Pop().empty());
DeviceSplitCandidate split;
split.loss_chg = 1.0f;
GPUExpandEntry root(0, 0, split, .0f, .0f, .0f);
split.left_sum = {0.0f, 1.0f};
split.right_sum = {0.0f, 1.0f};
GPUExpandEntry root(0, 0, split, 2.0f, 1.0f, 1.0f);
driver.Push({root});
EXPECT_EQ(driver.Pop().front().nid, 0);
driver.Push({GPUExpandEntry{1, 1, split, .0f, .0f, .0f}});
driver.Push({GPUExpandEntry{2, 1, split, .0f, .0f, .0f}});
driver.Push({GPUExpandEntry{3, 2, split, .0f, .0f, .0f}});
// Should return entries from level 1
driver.Push({GPUExpandEntry{1, 1, split, 2.0f, 1.0f, 1.0f}});
driver.Push({GPUExpandEntry{2, 1, split, 2.0f, 1.0f, 1.0f}});
driver.Push({GPUExpandEntry{3, 1, split, 2.0f, 1.0f, 1.0f}});
driver.Push({GPUExpandEntry{4, 2, split, 2.0f, 1.0f, 1.0f}});
// Should return 2 entries from level 1
// as we limited the driver to pop maximum 2 nodes
auto res = driver.Pop();
EXPECT_EQ(res.size(), 2);
for (auto &e : res) {
EXPECT_EQ(e.depth, 1);
}
// Should now return 1 entry from level 1
res = driver.Pop();
EXPECT_EQ(res[0].depth, 2);
EXPECT_EQ(res.size(), 1);
EXPECT_EQ(res.at(0).depth, 1);
res = driver.Pop();
EXPECT_EQ(res.at(0).depth, 2);
EXPECT_TRUE(driver.Pop().empty());
}
TEST(GpuHist, DriverLossGuided) {
DeviceSplitCandidate high_gain;
high_gain.left_sum = {0.0f, 1.0f};
high_gain.right_sum = {0.0f, 1.0f};
high_gain.loss_chg = 5.0f;
DeviceSplitCandidate low_gain;
DeviceSplitCandidate low_gain = high_gain;
low_gain.loss_chg = 1.0f;
Driver<GPUExpandEntry> driver(TrainParam::kLossGuide);
TrainParam p;
p.grow_policy=TrainParam::kLossGuide;
Driver<GPUExpandEntry> driver(p);
EXPECT_TRUE(driver.Pop().empty());
GPUExpandEntry root(0, 0, high_gain, .0f, .0f, .0f);
GPUExpandEntry root(0, 0, high_gain, 2.0f, 1.0f, 1.0f );
driver.Push({root});
EXPECT_EQ(driver.Pop().front().nid, 0);
// Select high gain first
driver.Push({GPUExpandEntry{1, 1, low_gain, .0f, .0f, .0f}});
driver.Push({GPUExpandEntry{2, 2, high_gain, .0f, .0f, .0f}});
driver.Push({GPUExpandEntry{1, 1, low_gain, 2.0f, 1.0f, 1.0f}});
driver.Push({GPUExpandEntry{2, 2, high_gain, 2.0f, 1.0f, 1.0f}});
auto res = driver.Pop();
EXPECT_EQ(res.size(), 1);
EXPECT_EQ(res[0].nid, 2);
@@ -49,8 +66,8 @@ TEST(GpuHist, DriverLossGuided) {
EXPECT_EQ(res[0].nid, 1);
// If equal gain, use nid
driver.Push({GPUExpandEntry{2, 1, low_gain, .0f, .0f, .0f}});
driver.Push({GPUExpandEntry{1, 1, low_gain, .0f, .0f, .0f}});
driver.Push({GPUExpandEntry{2, 1, low_gain, 2.0f, 1.0f, 1.0f}});
driver.Push({GPUExpandEntry{1, 1, low_gain, 2.0f, 1.0f, 1.0f}});
res = driver.Pop();
EXPECT_EQ(res[0].nid, 1);
res = driver.Pop();

1
tests/cpp/tree/gpu_hist/test_histogram.cu
View File

@@ -95,7 +95,6 @@ TEST(Histogram, GPUDeterministic) {
std::vector<int> shm_sizes{48 * 1024, 64 * 1024, 160 * 1024};
for (bool is_dense : is_dense_array) {
for (int shm_size : shm_sizes) {
TestDeterministicHistogram<GradientPair>(is_dense, shm_size);
TestDeterministicHistogram<GradientPairPrecise>(is_dense, shm_size);
}
}

45
tests/cpp/tree/test_gpu_hist.cu
View File

@@ -27,31 +27,40 @@ TEST(GpuHist, DeviceHistogram) {
// Ensures that node allocates correctly after reaching `kStopGrowingSize`.
dh::safe_cuda(cudaSetDevice(0));
constexpr size_t kNBins = 128;
constexpr size_t kNNodes = 4;
constexpr int kNNodes = 4;
constexpr size_t kStopGrowing = kNNodes * kNBins * 2u;
DeviceHistogram<GradientPairPrecise, kStopGrowing> histogram;
DeviceHistogramStorage<GradientPairPrecise, kStopGrowing> histogram;
histogram.Init(0, kNBins);
for (size_t i = 0; i < kNNodes; ++i) {
histogram.AllocateHistogram(i);
for (int i = 0; i < kNNodes; ++i) {
histogram.AllocateHistograms({i});
}
histogram.Reset();
ASSERT_EQ(histogram.Data().size(), kStopGrowing);
// Use allocated memory but do not erase nidx_map.
for (size_t i = 0; i < kNNodes; ++i) {
histogram.AllocateHistogram(i);
for (int i = 0; i < kNNodes; ++i) {
histogram.AllocateHistograms({i});
}
for (size_t i = 0; i < kNNodes; ++i) {
for (int i = 0; i < kNNodes; ++i) {
ASSERT_TRUE(histogram.HistogramExists(i));
}
// Erase existing nidx_map.
for (size_t i = kNNodes; i < kNNodes * 2; ++i) {
histogram.AllocateHistogram(i);
}
for (size_t i = 0; i < kNNodes; ++i) {
ASSERT_FALSE(histogram.HistogramExists(i));
// Add two new nodes
histogram.AllocateHistograms({kNNodes});
histogram.AllocateHistograms({kNNodes + 1});
// Old cached nodes should still exist
for (int i = 0; i < kNNodes; ++i) {
ASSERT_TRUE(histogram.HistogramExists(i));
}
// Should be deleted
ASSERT_FALSE(histogram.HistogramExists(kNNodes));
// Most recent node should exist
ASSERT_TRUE(histogram.HistogramExists(kNNodes + 1));
// Add same node again - should fail
EXPECT_ANY_THROW(histogram.AllocateHistograms({kNNodes + 1}););
}
std::vector<GradientPairPrecise> GetHostHistGpair() {
@@ -96,9 +105,9 @@ void TestBuildHist(bool use_shared_memory_histograms) {
thrust::host_vector<common::CompressedByteT> h_gidx_buffer (page->gidx_buffer.HostVector());
maker.row_partitioner.reset(new RowPartitioner(0, kNRows));
maker.hist.AllocateHistogram(0);
maker.hist.AllocateHistograms({0});
maker.gpair = gpair.DeviceSpan();
maker.histogram_rounding = CreateRoundingFactor<GradientSumT>(maker.gpair);;
maker.histogram_rounding = CreateRoundingFactor<GradientSumT>(maker.gpair);
BuildGradientHistogram(
page->GetDeviceAccessor(0), maker.feature_groups->DeviceAccessor(0),
@@ -106,7 +115,7 @@ void TestBuildHist(bool use_shared_memory_histograms) {
maker.hist.GetNodeHistogram(0), maker.histogram_rounding,
!use_shared_memory_histograms);
DeviceHistogram<GradientSumT>& d_hist = maker.hist;
DeviceHistogramStorage<GradientSumT>& d_hist = maker.hist;
auto node_histogram = d_hist.GetNodeHistogram(0);
// d_hist.data stored in float, not gradient pair
@@ -129,12 +138,10 @@ void TestBuildHist(bool use_shared_memory_histograms) {
TEST(GpuHist, BuildHistGlobalMem) {
TestBuildHist<GradientPairPrecise>(false);
TestBuildHist<GradientPair>(false);
}
TEST(GpuHist, BuildHistSharedMem) {
TestBuildHist<GradientPairPrecise>(true);
TestBuildHist<GradientPair>(true);
}
HistogramCutsWrapper GetHostCutMatrix () {
@@ -198,7 +205,7 @@ TEST(GpuHist, EvaluateRootSplit) {
// Initialize GPUHistMakerDevice::hist
maker.hist.Init(0, (max_bins - 1) * kNCols);
maker.hist.AllocateHistogram(0);
maker.hist.AllocateHistograms({0});
// Each row of hist_gpair represents gpairs for one feature.
// Each entry represents a bin.
std::vector<GradientPairPrecise> hist_gpair = GetHostHistGpair();