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

[EM] Avoid synchronous calls and unnecessary ATS access. (#10811)

Browse Source 
- Pass context into various functions.
- Factor out some CUDA algorithms.
- Use ATS only for update position.
This commit is contained in:
Jiaming Yuan
2024-09-10 14:33:14 +08:00
committed by GitHub
parent ed5f33df16
commit d94f6679fc
16 changed files with 161 additions and 201 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
tests/cpp/common/test_threading_utils.cu
View File

@@ -1,16 +1,17 @@
/**
* Copyright 2021-2023 by XGBoost Contributors
* Copyright 2021-2024, XGBoost Contributors
*/
#include <gtest/gtest.h>
#include <thrust/copy.h> // thrust::copy
#include "../../../src/common/device_helpers.cuh"
#include "../../../src/common/threading_utils.cuh"
#include "../helpers.h" // for MakeCUDACtx
namespace xgboost {
namespace common {
namespace xgboost::common {
TEST(SegmentedTrapezoidThreads, Basic) {
size_t constexpr kElements = 24, kGroups = 3;
auto ctx = MakeCUDACtx(0);
dh::device_vector<size_t> offset_ptr(kGroups + 1, 0);
offset_ptr[0] = 0;
offset_ptr[1] = 8;
@@ -19,11 +20,11 @@ TEST(SegmentedTrapezoidThreads, Basic) {
size_t h = 1;
dh::device_vector<size_t> thread_ptr(kGroups + 1, 0);
size_t total = SegmentedTrapezoidThreads(dh::ToSpan(offset_ptr), dh::ToSpan(thread_ptr), h);
size_t total = SegmentedTrapezoidThreads(&ctx, dh::ToSpan(offset_ptr), dh::ToSpan(thread_ptr), h);
ASSERT_EQ(total, kElements - kGroups);
h = 2;
SegmentedTrapezoidThreads(dh::ToSpan(offset_ptr), dh::ToSpan(thread_ptr), h);
SegmentedTrapezoidThreads(&ctx, dh::ToSpan(offset_ptr), dh::ToSpan(thread_ptr), h);
std::vector<size_t> h_thread_ptr(thread_ptr.size());
thrust::copy(thread_ptr.cbegin(), thread_ptr.cend(), h_thread_ptr.begin());
for (size_t i = 1; i < h_thread_ptr.size(); ++i) {
@@ -31,7 +32,7 @@ TEST(SegmentedTrapezoidThreads, Basic) {
}
h = 7;
SegmentedTrapezoidThreads(dh::ToSpan(offset_ptr), dh::ToSpan(thread_ptr), h);
SegmentedTrapezoidThreads(&ctx, dh::ToSpan(offset_ptr), dh::ToSpan(thread_ptr), h);
thrust::copy(thread_ptr.cbegin(), thread_ptr.cend(), h_thread_ptr.begin());
for (size_t i = 1; i < h_thread_ptr.size(); ++i) {
ASSERT_EQ(h_thread_ptr[i] - h_thread_ptr[i - 1], 28);
@@ -66,5 +67,4 @@ TEST(SegmentedTrapezoidThreads, Unravel) {
ASSERT_EQ(i, 6);
ASSERT_EQ(j, 7);
}
} // namespace common
} // namespace xgboost
} // namespace xgboost::common

23
tests/cpp/tree/gpu_hist/test_evaluate_splits.cu
View File

@@ -60,8 +60,7 @@ TEST_F(TestCategoricalSplitWithMissing, GPUHistEvaluator) {
GPUHistEvaluator evaluator{param_, static_cast<bst_feature_t>(feature_set.size()), ctx.Device()};
evaluator.Reset(cuts_, dh::ToSpan(feature_types), feature_set.size(), param_, false,
ctx.Device());
evaluator.Reset(&ctx, cuts_, dh::ToSpan(feature_types), feature_set.size(), param_, false);
DeviceSplitCandidate result = evaluator.EvaluateSingleSplit(&ctx, input, shared_inputs).split;
ASSERT_EQ(result.thresh, 1);
@@ -104,7 +103,7 @@ TEST(GpuHist, PartitionBasic) {
};
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_set.size()), ctx.Device()};
evaluator.Reset(cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, false, ctx.Device());
evaluator.Reset(&ctx, cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, false);
{
// -1.0s go right
@@ -217,7 +216,7 @@ TEST(GpuHist, PartitionTwoFeatures) {
false};
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_set.size()), ctx.Device()};
evaluator.Reset(cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, false, ctx.Device());
evaluator.Reset(&ctx, cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, false);
{
auto parent_sum = quantiser.ToFixedPoint(GradientPairPrecise{-6.0, 3.0});
@@ -277,10 +276,8 @@ TEST(GpuHist, PartitionTwoNodes) {
cuts.min_vals_.ConstDeviceSpan(),
false};
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_set.size()),
ctx.Device()};
evaluator.Reset(cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, false,
ctx.Device());
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_set.size()), ctx.Device()};
evaluator.Reset(&ctx, cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, false);
{
auto parent_sum = quantiser.ToFixedPoint(GradientPairPrecise{-6.0, 3.0});
@@ -336,10 +333,8 @@ void TestEvaluateSingleSplit(bool is_categorical) {
cuts.min_vals_.ConstDeviceSpan(),
false};
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_set.size()),
ctx.Device()};
evaluator.Reset(cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, false,
ctx.Device());
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_set.size()), ctx.Device()};
evaluator.Reset(&ctx, cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, false);
DeviceSplitCandidate result = evaluator.EvaluateSingleSplit(&ctx, input, shared_inputs).split;
EXPECT_EQ(result.findex, 1);
@@ -522,7 +517,7 @@ TEST_F(TestPartitionBasedSplit, GpuHist) {
cuts_.cut_values_.SetDevice(ctx.Device());
cuts_.min_vals_.SetDevice(ctx.Device());
evaluator.Reset(cuts_, dh::ToSpan(ft), info_.num_col_, param_, false, ctx.Device());
evaluator.Reset(&ctx, cuts_, dh::ToSpan(ft), info_.num_col_, param_, false);
// Convert the sample histogram to fixed point
auto quantiser = DummyRoundingFactor(&ctx);
@@ -586,7 +581,7 @@ void VerifyColumnSplitEvaluateSingleSplit(bool is_categorical) {
false};
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_set.size()), ctx.Device()};
evaluator.Reset(cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, true, ctx.Device());
evaluator.Reset(&ctx, cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, true);
DeviceSplitCandidate result = evaluator.EvaluateSingleSplit(&ctx, input, shared_inputs).split;
EXPECT_EQ(result.findex, 1);