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

Use quantised gradients in gpu_hist histograms (#8246)

Browse Source
This commit is contained in:
Rory Mitchell
2022-09-26 17:35:35 +02:00
committed by GitHub
parent 4056974e37
commit 8f77677193
14 changed files with 394 additions and 336 deletions
Download Patch File Download Diff File Expand all files Collapse all files

132
tests/cpp/tree/gpu_hist/test_histogram.cu
View File

@@ -10,7 +10,6 @@
namespace xgboost {
namespace tree {
template <typename Gradient>
void TestDeterministicHistogram(bool is_dense, int shm_size) {
size_t constexpr kBins = 256, kCols = 120, kRows = 16384, kRounds = 16;
float constexpr kLower = -1e-2, kUpper = 1e2;
@@ -26,41 +25,41 @@ void TestDeterministicHistogram(bool is_dense, int shm_size) {
auto ridx = row_partitioner.GetRows(0);
int num_bins = kBins * kCols;
dh::device_vector<Gradient> histogram(num_bins);
dh::device_vector<GradientPairInt64> histogram(num_bins);
auto d_histogram = dh::ToSpan(histogram);
auto gpair = GenerateRandomGradients(kRows, kLower, kUpper);
gpair.SetDevice(0);
FeatureGroups feature_groups(page->Cuts(), page->is_dense, shm_size,
sizeof(Gradient));
sizeof(GradientPairInt64));
auto rounding = CreateRoundingFactor<Gradient>(gpair.DeviceSpan());
auto rounding = GradientQuantizer(gpair.DeviceSpan());
BuildGradientHistogram(page->GetDeviceAccessor(0),
feature_groups.DeviceAccessor(0), gpair.DeviceSpan(),
ridx, d_histogram, rounding);
std::vector<Gradient> histogram_h(num_bins);
std::vector<GradientPairInt64> histogram_h(num_bins);
dh::safe_cuda(cudaMemcpy(histogram_h.data(), d_histogram.data(),
num_bins * sizeof(Gradient),
num_bins * sizeof(GradientPairInt64),
cudaMemcpyDeviceToHost));
for (size_t i = 0; i < kRounds; ++i) {
dh::device_vector<Gradient> new_histogram(num_bins);
dh::device_vector<GradientPairInt64> new_histogram(num_bins);
auto d_new_histogram = dh::ToSpan(new_histogram);
auto rounding = CreateRoundingFactor<Gradient>(gpair.DeviceSpan());
auto rounding = GradientQuantizer(gpair.DeviceSpan());
BuildGradientHistogram(page->GetDeviceAccessor(0),
feature_groups.DeviceAccessor(0),
gpair.DeviceSpan(), ridx, d_new_histogram,
rounding);
std::vector<Gradient> new_histogram_h(num_bins);
std::vector<GradientPairInt64> new_histogram_h(num_bins);
dh::safe_cuda(cudaMemcpy(new_histogram_h.data(), d_new_histogram.data(),
num_bins * sizeof(Gradient),
num_bins * sizeof(GradientPairInt64),
cudaMemcpyDeviceToHost));
for (size_t j = 0; j < new_histogram_h.size(); ++j) {
ASSERT_EQ(new_histogram_h[j].GetGrad(), histogram_h[j].GetGrad());
ASSERT_EQ(new_histogram_h[j].GetHess(), histogram_h[j].GetHess());
ASSERT_EQ(new_histogram_h[j].GetQuantisedGrad(), histogram_h[j].GetQuantisedGrad());
ASSERT_EQ(new_histogram_h[j].GetQuantisedHess(), histogram_h[j].GetQuantisedHess());
}
}
@@ -71,20 +70,20 @@ void TestDeterministicHistogram(bool is_dense, int shm_size) {
// Use a single feature group to compute the baseline.
FeatureGroups single_group(page->Cuts());
dh::device_vector<Gradient> baseline(num_bins);
dh::device_vector<GradientPairInt64> baseline(num_bins);
BuildGradientHistogram(page->GetDeviceAccessor(0),
single_group.DeviceAccessor(0),
gpair.DeviceSpan(), ridx, dh::ToSpan(baseline),
rounding);
std::vector<Gradient> baseline_h(num_bins);
std::vector<GradientPairInt64> baseline_h(num_bins);
dh::safe_cuda(cudaMemcpy(baseline_h.data(), baseline.data().get(),
num_bins * sizeof(Gradient),
num_bins * sizeof(GradientPairInt64),
cudaMemcpyDeviceToHost));
for (size_t i = 0; i < baseline.size(); ++i) {
EXPECT_NEAR(baseline_h[i].GetGrad(), histogram_h[i].GetGrad(),
baseline_h[i].GetGrad() * 1e-3);
EXPECT_NEAR(baseline_h[i].GetQuantisedGrad(), histogram_h[i].GetQuantisedGrad(),
baseline_h[i].GetQuantisedGrad() * 1e-3);
}
}
}
@@ -95,11 +94,25 @@ 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<GradientPairPrecise>(is_dense, shm_size);
TestDeterministicHistogram(is_dense, shm_size);
}
}
}
void ValidateCategoricalHistogram(size_t n_categories, common::Span<GradientPairInt64> onehot,
common::Span<GradientPairInt64> cat) {
auto cat_sum = std::accumulate(cat.cbegin(), cat.cend(), GradientPairInt64{});
for (size_t c = 0; c < n_categories; ++c) {
auto zero = onehot[c * 2];
auto one = onehot[c * 2 + 1];
auto chosen = cat[c];
auto not_chosen = cat_sum - chosen;
ASSERT_EQ(zero, not_chosen);
ASSERT_EQ(one, chosen);
}
}
// Test 1 vs rest categorical histogram is equivalent to one hot encoded data.
void TestGPUHistogramCategorical(size_t num_categories) {
size_t constexpr kRows = 340;
@@ -110,10 +123,10 @@ void TestGPUHistogramCategorical(size_t num_categories) {
BatchParam batch_param{0, static_cast<int32_t>(kBins)};
tree::RowPartitioner row_partitioner(0, kRows);
auto ridx = row_partitioner.GetRows(0);
dh::device_vector<GradientPairPrecise> cat_hist(num_categories);
dh::device_vector<GradientPairInt64> cat_hist(num_categories);
auto gpair = GenerateRandomGradients(kRows, 0, 2);
gpair.SetDevice(0);
auto rounding = CreateRoundingFactor<GradientPairPrecise>(gpair.DeviceSpan());
auto rounding = GradientQuantizer(gpair.DeviceSpan());
/**
* Generate hist with cat data.
*/
@@ -131,7 +144,7 @@ void TestGPUHistogramCategorical(size_t num_categories) {
*/
auto x_encoded = OneHotEncodeFeature(x, num_categories);
auto encode_m = GetDMatrixFromData(x_encoded, kRows, num_categories);
dh::device_vector<GradientPairPrecise> encode_hist(2 * num_categories);
dh::device_vector<GradientPairInt64> encode_hist(2 * num_categories);
for (auto const &batch : encode_m->GetBatches<EllpackPage>(batch_param)) {
auto* page = batch.Impl();
FeatureGroups single_group(page->Cuts());
@@ -141,14 +154,14 @@ void TestGPUHistogramCategorical(size_t num_categories) {
rounding);
}
std::vector<GradientPairPrecise> h_cat_hist(cat_hist.size());
std::vector<GradientPairInt64> h_cat_hist(cat_hist.size());
thrust::copy(cat_hist.begin(), cat_hist.end(), h_cat_hist.begin());
std::vector<GradientPairPrecise> h_encode_hist(encode_hist.size());
std::vector<GradientPairInt64> h_encode_hist(encode_hist.size());
thrust::copy(encode_hist.begin(), encode_hist.end(), h_encode_hist.begin());
ValidateCategoricalHistogram(num_categories,
common::Span<GradientPairPrecise>{h_encode_hist},
common::Span<GradientPairPrecise>{h_cat_hist});
common::Span<GradientPairInt64>{h_encode_hist},
common::Span<GradientPairInt64>{h_cat_hist});
}
TEST(Histogram, GPUHistCategorical) {
@@ -156,5 +169,74 @@ TEST(Histogram, GPUHistCategorical) {
TestGPUHistogramCategorical(num_categories);
}
}
namespace {
// Atomic add as type cast for test.
XGBOOST_DEV_INLINE int64_t atomicAdd(int64_t *dst, int64_t src) { // NOLINT
uint64_t* u_dst = reinterpret_cast<uint64_t*>(dst);
uint64_t u_src = *reinterpret_cast<uint64_t*>(&src);
uint64_t ret = ::atomicAdd(u_dst, u_src);
return *reinterpret_cast<int64_t*>(&ret);
}
}
void TestAtomicAdd() {
size_t n_elements = 1024;
dh::device_vector<int64_t> result_a(1, 0);
auto d_result_a = result_a.data().get();
dh::device_vector<int64_t> result_b(1, 0);
auto d_result_b = result_b.data().get();
/**
* Test for simple inputs
*/
std::vector<int64_t> h_inputs(n_elements);
for (size_t i = 0; i < h_inputs.size(); ++i) {
h_inputs[i] = (i % 2 == 0) ? i : -i;
}
dh::device_vector<int64_t> inputs(h_inputs);
auto d_inputs = inputs.data().get();
dh::LaunchN(n_elements, [=] __device__(size_t i) {
AtomicAdd64As32(d_result_a, d_inputs[i]);
atomicAdd(d_result_b, d_inputs[i]);
});
ASSERT_EQ(result_a[0], result_b[0]);
/**
* Test for positive values that don't fit into 32 bit integer.
*/
thrust::fill(inputs.begin(), inputs.end(),
(std::numeric_limits<uint32_t>::max() / 2));
thrust::fill(result_a.begin(), result_a.end(), 0);
thrust::fill(result_b.begin(), result_b.end(), 0);
dh::LaunchN(n_elements, [=] __device__(size_t i) {
AtomicAdd64As32(d_result_a, d_inputs[i]);
atomicAdd(d_result_b, d_inputs[i]);
});
ASSERT_EQ(result_a[0], result_b[0]);
ASSERT_GT(result_a[0], std::numeric_limits<uint32_t>::max());
CHECK_EQ(thrust::reduce(inputs.begin(), inputs.end(), int64_t(0)), result_a[0]);
/**
* Test for negative values that don't fit into 32 bit integer.
*/
thrust::fill(inputs.begin(), inputs.end(),
(std::numeric_limits<int32_t>::min() / 2));
thrust::fill(result_a.begin(), result_a.end(), 0);
thrust::fill(result_b.begin(), result_b.end(), 0);
dh::LaunchN(n_elements, [=] __device__(size_t i) {
AtomicAdd64As32(d_result_a, d_inputs[i]);
atomicAdd(d_result_b, d_inputs[i]);
});
ASSERT_EQ(result_a[0], result_b[0]);
ASSERT_LT(result_a[0], std::numeric_limits<int32_t>::min());
CHECK_EQ(thrust::reduce(inputs.begin(), inputs.end(), int64_t(0)), result_a[0]);
}
TEST(Histogram, AtomicAddInt64) {
TestAtomicAdd();
}
} // namespace tree
} // namespace xgboost