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

[breaking] Use integer atomic for GPU histogram. (#7180)

Browse Source 
On GPU we use rouding factor to truncate the gradient for deterministic results. This PR changes the gradient representation to fixed point number with exponent aligned with rounding factor.

    [breaking] Drop non-deterministic histogram.
    Use fixed point for shared memory.

This PR is to improve the performance of GPU Hist. 

Co-authored-by: Andy Adinets <aadinets@nvidia.com>
This commit is contained in:
Jiaming Yuan
2021-08-28 05:17:05 +08:00
committed by GitHub
parent e7d7ab6bc3
commit 7a1d67f9cb
11 changed files with 295 additions and 142 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
tests/cpp/tree/test_gpu_hist.cu
View File

@@ -1,8 +1,9 @@
/*!
* Copyright 2017-2020 XGBoost contributors
* Copyright 2017-2021 XGBoost contributors
*/
#include <gtest/gtest.h>
#include <thrust/device_vector.h>
#include <thrust/host_vector.h>
#include <dmlc/filesystem.h>
#include <xgboost/base.h>
#include <random>
@@ -80,8 +81,8 @@ void TestBuildHist(bool use_shared_memory_histograms) {
param.Init(args);
auto page = BuildEllpackPage(kNRows, kNCols);
BatchParam batch_param{};
GPUHistMakerDevice<GradientSumT> maker(0, page.get(), {}, kNRows, param, kNCols, kNCols,
true, batch_param);
GPUHistMakerDevice<GradientSumT> maker(0, page.get(), {}, kNRows, param,
kNCols, kNCols, batch_param);
xgboost::SimpleLCG gen;
xgboost::SimpleRealUniformDistribution<bst_float> dist(0.0f, 1.0f);
HostDeviceVector<GradientPair> gpair(kNRows);
@@ -93,14 +94,18 @@ void TestBuildHist(bool use_shared_memory_histograms) {
gpair.SetDevice(0);
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.gpair = gpair.DeviceSpan();
maker.histogram_rounding = CreateRoundingFactor<GradientSumT>(maker.gpair);;
maker.BuildHist(0);
DeviceHistogram<GradientSumT> d_hist = maker.hist;
BuildGradientHistogram(
page->GetDeviceAccessor(0), maker.feature_groups->DeviceAccessor(0),
gpair.DeviceSpan(), maker.row_partitioner->GetRows(0),
maker.hist.GetNodeHistogram(0), maker.histogram_rounding,
!use_shared_memory_histograms);
DeviceHistogram<GradientSumT>& d_hist = maker.hist;
auto node_histogram = d_hist.GetNodeHistogram(0);
// d_hist.data stored in float, not gradient pair
@@ -115,6 +120,7 @@ void TestBuildHist(bool use_shared_memory_histograms) {
std::vector<GradientPairPrecise> solution = GetHostHistGpair();
std::cout << std::fixed;
for (size_t i = 0; i < h_result.size(); ++i) {
ASSERT_FALSE(std::isnan(h_result[i].GetGrad()));
EXPECT_NEAR(h_result[i].GetGrad(), solution[i].GetGrad(), 0.01f);
EXPECT_NEAR(h_result[i].GetHess(), solution[i].GetHess(), 0.01f);
}
@@ -158,7 +164,8 @@ TEST(GpuHist, ApplySplit) {
HostDeviceVector<FeatureType> feature_types(10, FeatureType::kCategorical);
feature_types.SetDevice(bparam.gpu_id);
tree::GPUHistMakerDevice<GradientPairPrecise> updater(
0, impl, feature_types.ConstDeviceSpan(), n_rows, tparam, 0, n_cols, true, bparam);
0, impl, feature_types.ConstDeviceSpan(), n_rows, tparam, 0, n_cols,
bparam);
updater.ApplySplit(candidate, &tree);
ASSERT_EQ(tree.GetSplitTypes().size(), 3);
@@ -217,8 +224,8 @@ TEST(GpuHist, EvaluateRootSplit) {
// Initialize GPUHistMakerDevice
auto page = BuildEllpackPage(kNRows, kNCols);
BatchParam batch_param{};
GPUHistMakerDevice<GradientPairPrecise>
maker(0, page.get(), {}, kNRows, param, kNCols, kNCols, true, batch_param);
GPUHistMakerDevice<GradientPairPrecise> maker(
0, page.get(), {}, kNRows, param, kNCols, kNCols, batch_param);
// Initialize GPUHistMakerDevice::node_sum_gradients
maker.node_sum_gradients = {};