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

Serialize expand entry for allgather. (#9702)

Browse Source
This commit is contained in:
Jiaming Yuan 2023-10-24 14:33:28 +08:00 committed by GitHub
parent ee8b29c843
commit 7a02facc9d
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
14 changed files with 336 additions and 76 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
include/xgboost/json.h
View File

@ -153,7 +153,7 @@ class JsonTypedArray : public Value {
using Type = T;
JsonTypedArray() : Value(kind) {}
explicit JsonTypedArray(size_t n) : Value(kind) { vec_.resize(n); }
explicit JsonTypedArray(std::size_t n) : Value(kind) { vec_.resize(n); }
JsonTypedArray(JsonTypedArray&& that) noexcept : Value{kind}, vec_{std::move(that.vec_)} {}
bool operator==(Value const& rhs) const override;
@ -171,21 +171,21 @@ class JsonTypedArray : public Value {
};
/**
* \brief Typed UBJSON array for 32-bit floating point.
* @brief Typed UBJSON array for 32-bit floating point.
*/
using F32Array = JsonTypedArray<float, Value::ValueKind::kNumberArray>;
/**
* \brief Typed UBJSON array for uint8_t.
* @brief Typed UBJSON array for uint8_t.
*/
using U8Array = JsonTypedArray<uint8_t, Value::ValueKind::kU8Array>;
using U8Array = JsonTypedArray<std::uint8_t, Value::ValueKind::kU8Array>;
/**
* \brief Typed UBJSON array for int32_t.
* @brief Typed UBJSON array for int32_t.
*/
using I32Array = JsonTypedArray<int32_t, Value::ValueKind::kI32Array>;
using I32Array = JsonTypedArray<std::int32_t, Value::ValueKind::kI32Array>;
/**
* \brief Typed UBJSON array for int64_t.
* @brief Typed UBJSON array for int64_t.
*/
using I64Array = JsonTypedArray<int64_t, Value::ValueKind::kI64Array>;
using I64Array = JsonTypedArray<std::int64_t, Value::ValueKind::kI64Array>;
class JsonObject : public Value {
public:

9
src/collective/allgather.h
View File

@ -9,7 +9,8 @@
#include <type_traits> // for remove_cv_t
#include <vector> // for vector
#include "comm.h" // for Comm, Channel, EraseType
#include "../common/type.h" // for EraseType
#include "comm.h" // for Comm, Channel
#include "xgboost/collective/result.h" // for Result
#include "xgboost/span.h" // for Span
@ -33,7 +34,7 @@ namespace cpu_impl {
template <typename T>
[[nodiscard]] Result RingAllgather(Comm const& comm, common::Span<T> data, std::size_t size) {
auto n_bytes = sizeof(T) * size;
auto erased = EraseType(data);
auto erased = common::EraseType(data);
auto rank = comm.Rank();
auto prev = BootstrapPrev(rank, comm.World());
@ -65,8 +66,8 @@ template <typename T>
auto n_total_bytes = std::accumulate(sizes.cbegin(), sizes.cend(), 0);
result.resize(n_total_bytes / sizeof(T));
auto h_result = common::Span{result.data(), result.size()};
auto erased_result = EraseType(h_result);
auto erased_data = EraseType(data);
auto erased_result = common::EraseType(h_result);
auto erased_data = common::EraseType(data);
std::vector<std::int64_t> offset(world + 1);
return cpu_impl::RingAllgatherV(comm, sizes, erased_data,

9
src/collective/allreduce.h
View File

@ -4,8 +4,9 @@
#pragma once
#include <cstdint> // for int8_t
#include <functional> // for function
#include <type_traits> // for is_invocable_v
#include <type_traits> // for is_invocable_v, enable_if_t
#include "../common/type.h" // for EraseType, RestoreType
#include "../data/array_interface.h" // for ArrayInterfaceHandler
#include "comm.h" // for Comm, RestoreType
#include "xgboost/collective/result.h" // for Result
@ -23,14 +24,14 @@ Result RingAllreduce(Comm const& comm, common::Span<std::int8_t> data, Func cons
template <typename T, typename Fn>
std::enable_if_t<std::is_invocable_v<Fn, common::Span<T const>, common::Span<T>>, Result> Allreduce(
Comm const& comm, common::Span<T> data, Fn redop) {
auto erased = EraseType(data);
auto erased = common::EraseType(data);
auto type = ToDType<T>::kType;
auto erased_fn = [type, redop](common::Span<std::int8_t const> lhs,
common::Span<std::int8_t> out) {
CHECK_EQ(lhs.size(), out.size()) << "Invalid input for reduction.";
auto lhs_t = RestoreType<T const>(lhs);
auto rhs_t = RestoreType<T>(out);
auto lhs_t = common::RestoreType<T const>(lhs);
auto rhs_t = common::RestoreType<T>(out);
redop(lhs_t, rhs_t);
};

16
src/collective/comm.h
View File

@ -137,20 +137,4 @@ class Channel {
};
enum class Op { kMax = 0, kMin = 1, kSum = 2, kBitwiseAND = 3, kBitwiseOR = 4, kBitwiseXOR = 5 };
template <typename T, typename U = std::conditional_t<std::is_const_v<T>,
std::add_const_t<std::int8_t>, std::int8_t>>
common::Span<U> EraseType(common::Span<T> data) {
auto n_total_bytes = data.size_bytes();
auto erased = common::Span{reinterpret_cast<std::add_pointer_t<U>>(data.data()), n_total_bytes};
return erased;
}
template <typename T, typename U>
common::Span<T> RestoreType(common::Span<U> data) {
static_assert(std::is_same_v<std::remove_const_t<U>, std::int8_t>);
auto n_total_bytes = data.size_bytes();
auto restored = common::Span{reinterpret_cast<T*>(data.data()), n_total_bytes / sizeof(T)};
return restored;
}
} // namespace xgboost::collective

1
src/collective/communicator-inl.h
View File

@ -327,6 +327,5 @@ inline SpecialAllgatherVResult<T> SpecialAllgatherV(std::vector<T> const &inputs
return {offsets, all_sizes, all_inputs};
}
} // namespace collective
} // namespace xgboost

2
src/common/bitfield.h
View File

@ -9,7 +9,7 @@
#include <bitset> // for bitset
#include <cstdint> // for uint32_t, uint64_t, uint8_t
#include <ostream> // for ostream
#include <type_traits> // for conditional_t, is_signed_v
#include <type_traits> // for conditional_t, is_signed_v, add_const_t
#if defined(__CUDACC__)
#include <thrust/copy.h>

24
src/common/type.h Normal file
View File

@ -0,0 +1,24 @@
/**
* Copyright 2023, XGBoost Contributors
*/
#pragma once
#include <cstdint> // for int8_t
#include <type_traits> // for is_const_v, add_const_t, conditional_t, add_pointer_t
#include "xgboost/span.h" // for Span
namespace xgboost::common {
template <typename T, typename U = std::conditional_t<std::is_const_v<T>,
std::add_const_t<std::int8_t>, std::int8_t>>
common::Span<U> EraseType(common::Span<T> data) {
auto n_total_bytes = data.size_bytes();
auto erased = common::Span{reinterpret_cast<std::add_pointer_t<U>>(data.data()), n_total_bytes};
return erased;
}
template <typename T, typename U>
common::Span<T> RestoreType(common::Span<U> data) {
auto n_total_bytes = data.size_bytes();
auto restored = common::Span{reinterpret_cast<T*>(data.data()), n_total_bytes / sizeof(T)};
return restored;
}
} // namespace xgboost::common

115
src/tree/gpu_hist/expand_entry.cuh
View File

@ -1,31 +1,36 @@
/*!
* Copyright 2020 by XGBoost Contributors
/**
* Copyright 2020-2023, XGBoost Contributors
*/
#ifndef EXPAND_ENTRY_CUH_
#define EXPAND_ENTRY_CUH_
#include <xgboost/span.h>
#include <limits> // for numeric_limits
#include <utility> // for move
#include "../param.h"
#include "../updater_gpu_common.cuh"
#include "xgboost/base.h" // for bst_node_t
namespace xgboost {
namespace tree {
namespace xgboost::tree {
struct GPUExpandEntry {
int nid;
int depth;
bst_node_t nid;
bst_node_t depth;
DeviceSplitCandidate split;
float base_weight { std::numeric_limits<float>::quiet_NaN() };
float left_weight { std::numeric_limits<float>::quiet_NaN() };
float right_weight { std::numeric_limits<float>::quiet_NaN() };
float base_weight{std::numeric_limits<float>::quiet_NaN()};
float left_weight{std::numeric_limits<float>::quiet_NaN()};
float right_weight{std::numeric_limits<float>::quiet_NaN()};
GPUExpandEntry() = default;
XGBOOST_DEVICE GPUExpandEntry(int nid, int depth, DeviceSplitCandidate split,
float base, float left, float right)
: nid(nid), depth(depth), split(std::move(split)), base_weight{base},
left_weight{left}, right_weight{right} {}
bool IsValid(const TrainParam& param, int num_leaves) const {
XGBOOST_DEVICE GPUExpandEntry(bst_node_t nid, bst_node_t depth, DeviceSplitCandidate split,
float base, float left, float right)
: nid(nid),
depth(depth),
split(std::move(split)),
base_weight{base},
left_weight{left},
right_weight{right} {}
[[nodiscard]] bool IsValid(TrainParam const& param, bst_node_t num_leaves) const {
if (split.loss_chg <= kRtEps) return false;
if (split.left_sum.GetQuantisedHess() == 0 || split.right_sum.GetQuantisedHess() == 0) {
return false;
@ -42,17 +47,11 @@ struct GPUExpandEntry {
return true;
}
bst_float GetLossChange() const {
return split.loss_chg;
}
[[nodiscard]] float GetLossChange() const { return split.loss_chg; }
int GetNodeId() const {
return nid;
}
[[nodiscard]] bst_node_t GetNodeId() const { return nid; }
int GetDepth() const {
return depth;
}
[[nodiscard]] bst_node_t GetDepth() const { return depth; }
friend std::ostream& operator<<(std::ostream& os, const GPUExpandEntry& e) {
os << "GPUExpandEntry: \n";
@ -63,9 +62,69 @@ struct GPUExpandEntry {
os << "right_sum: " << e.split.right_sum << "\n";
return os;
}
};
} // namespace tree
} // namespace xgboost
void Save(Json* p_out) const {
auto& out = *p_out;
out["nid"] = Integer{this->nid};
out["depth"] = Integer{this->depth};
// GPU specific
out["base_weight"] = this->base_weight;
out["left_weight"] = this->left_weight;
out["right_weight"] = this->right_weight;
/**
* Handle split
*/
out["split"] = Object{};
auto& split = out["split"];
split["loss_chg"] = this->split.loss_chg;
split["sindex"] = Integer{this->split.findex};
split["split_value"] = this->split.fvalue;
// cat
split["thresh"] = Integer{this->split.thresh};
split["is_cat"] = Boolean{this->split.is_cat};
/**
* Gradients
*/
auto save = [&](std::string const& name, GradientPairInt64 const& sum) {
out[name] = I64Array{2};
auto& array = get<I64Array>(out[name]);
array[0] = sum.GetQuantisedGrad();
array[1] = sum.GetQuantisedHess();
};
save("left_sum", this->split.left_sum);
save("right_sum", this->split.right_sum);
}
void Load(Json const& in) {
this->nid = get<Integer const>(in["nid"]);
this->depth = get<Integer const>(in["depth"]);
// GPU specific
this->base_weight = get<Number const>(in["base_weight"]);
this->left_weight = get<Number const>(in["left_weight"]);
this->right_weight = get<Number const>(in["right_weight"]);
/**
* Handle split
*/
auto const& split = in["split"];
this->split.loss_chg = get<Number const>(split["loss_chg"]);
this->split.findex = get<Integer const>(split["sindex"]);
this->split.fvalue = get<Number const>(split["split_value"]);
// cat
this->split.thresh = get<Integer const>(split["thresh"]);
this->split.is_cat = get<Boolean const>(split["is_cat"]);
/**
* Gradients
*/
auto const& left_sum = get<I64Array const>(in["left_sum"]);
this->split.left_sum = GradientPairInt64{left_sum[0], left_sum[1]};
auto const& right_sum = get<I64Array const>(in["right_sum"]);
this->split.right_sum = GradientPairInt64{right_sum[0], right_sum[1]};
}
};
} // namespace xgboost::tree
#endif // EXPAND_ENTRY_CUH_

122
src/tree/hist/expand_entry.h
View File

@ -1,16 +1,20 @@
/**
* Copyright 2021-2023 XGBoost contributors
* Copyright 2021-2023, XGBoost Contributors
*/
#ifndef XGBOOST_TREE_HIST_EXPAND_ENTRY_H_
#define XGBOOST_TREE_HIST_EXPAND_ENTRY_H_
#include <algorithm> // for all_of
#include <ostream> // for ostream
#include <utility> // for move
#include <vector> // for vector
#include <algorithm> // for all_of
#include <ostream> // for ostream
#include <string> // for string
#include <type_traits> // for add_const_t
#include <utility> // for move
#include <vector> // for vector
#include "../param.h" // for SplitEntry, SplitEntryContainer, TrainParam
#include "xgboost/base.h" // for GradientPairPrecise, bst_node_t
#include "../../common/type.h" // for EraseType
#include "../param.h" // for SplitEntry, SplitEntryContainer, TrainParam
#include "xgboost/base.h" // for GradientPairPrecise, bst_node_t
#include "xgboost/json.h" // for Json
namespace xgboost::tree {
/**
@ -29,6 +33,66 @@ struct ExpandEntryImpl {
[[nodiscard]] bool IsValid(TrainParam const& param, bst_node_t num_leaves) const {
return static_cast<Impl const*>(this)->IsValidImpl(param, num_leaves);
}
void Save(Json* p_out) const {
auto& out = *p_out;
auto self = static_cast<Impl const*>(this);
out["nid"] = Integer{this->nid};
out["depth"] = Integer{this->depth};
/**
* Handle split
*/
out["split"] = Object{};
auto& split = out["split"];
split["loss_chg"] = self->split.loss_chg;
split["sindex"] = Integer{self->split.sindex};
split["split_value"] = self->split.split_value;
auto const& cat_bits = self->split.cat_bits;
auto s_cat_bits = common::Span{cat_bits.data(), cat_bits.size()};
split["cat_bits"] = U8Array{s_cat_bits.size_bytes()};
auto& j_cat_bits = get<U8Array>(split["cat_bits"]);
using T = typename decltype(self->split.cat_bits)::value_type;
auto erased =
common::EraseType<std::add_const_t<T>, std::add_const_t<std::uint8_t>>(s_cat_bits);
for (std::size_t i = 0; i < erased.size(); ++i) {
j_cat_bits[i] = erased[i];
}
split["is_cat"] = Boolean{self->split.is_cat};
self->SaveGrad(&split);
}
void Load(Json const& in) {
auto self = static_cast<Impl*>(this);
this->nid = get<Integer const>(in["nid"]);
this->depth = get<Integer const>(in["depth"]);
/**
* Handle split
*/
auto const& split = in["split"];
self->split.loss_chg = get<Number const>(split["loss_chg"]);
self->split.sindex = get<Integer const>(split["sindex"]);
self->split.split_value = get<Number const>(split["split_value"]);
auto const& j_cat_bits = get<U8Array const>(split["cat_bits"]);
using T = typename decltype(self->split.cat_bits)::value_type;
auto restored = common::RestoreType<std::add_const_t<T>>(
common::Span{j_cat_bits.data(), j_cat_bits.size()});
self->split.cat_bits.resize(restored.size());
for (std::size_t i = 0; i < restored.size(); ++i) {
self->split.cat_bits[i] = restored[i];
}
self->split.is_cat = get<Boolean const>(split["is_cat"]);
self->LoadGrad(split);
}
};
struct CPUExpandEntry : public ExpandEntryImpl<CPUExpandEntry> {
@ -39,6 +103,24 @@ struct CPUExpandEntry : public ExpandEntryImpl<CPUExpandEntry> {
: ExpandEntryImpl{nidx, depth}, split(std::move(split)) {}
CPUExpandEntry(bst_node_t nidx, bst_node_t depth) : ExpandEntryImpl{nidx, depth} {}
void SaveGrad(Json* p_out) const {
auto& out = *p_out;
auto save = [&](std::string const& name, GradStats const& sum) {
out[name] = F32Array{2};
auto& array = get<F32Array>(out[name]);
array[0] = sum.GetGrad();
array[1] = sum.GetHess();
};
save("left_sum", this->split.left_sum);
save("right_sum", this->split.right_sum);
}
void LoadGrad(Json const& in) {
auto const& left_sum = get<F32Array const>(in["left_sum"]);
this->split.left_sum = GradStats{left_sum[0], left_sum[1]};
auto const& right_sum = get<F32Array const>(in["right_sum"]);
this->split.right_sum = GradStats{right_sum[0], right_sum[1]};
}
[[nodiscard]] bool IsValidImpl(TrainParam const& param, bst_node_t num_leaves) const {
if (split.loss_chg <= kRtEps) return false;
if (split.left_sum.GetHess() == 0 || split.right_sum.GetHess() == 0) {
@ -88,6 +170,32 @@ struct MultiExpandEntry : public ExpandEntryImpl<MultiExpandEntry> {
MultiExpandEntry() = default;
MultiExpandEntry(bst_node_t nidx, bst_node_t depth) : ExpandEntryImpl{nidx, depth} {}
void SaveGrad(Json* p_out) const {
auto& out = *p_out;
auto save = [&](std::string const& name, std::vector<GradientPairPrecise> const& sum) {
out[name] = F32Array{sum.size() * 2};
auto& array = get<F32Array>(out[name]);
for (std::size_t i = 0, j = 0; i < sum.size(); i++, j += 2) {
array[j] = sum[i].GetGrad();
array[j + 1] = sum[i].GetHess();
}
};
save("left_sum", this->split.left_sum);
save("right_sum", this->split.right_sum);
}
void LoadGrad(Json const& in) {
auto load = [&](std::string const& name, std::vector<GradientPairPrecise>* p_sum) {
auto const& array = get<F32Array const>(in[name]);
auto& sum = *p_sum;
sum.resize(array.size() / 2);
for (std::size_t i = 0, j = 0; i < sum.size(); ++i, j += 2) {
sum[i] = GradientPairPrecise{array[j], array[j + 1]};
}
};
load("left_sum", &this->split.left_sum);
load("right_sum", &this->split.right_sum);
}
[[nodiscard]] bool IsValidImpl(TrainParam const& param, bst_node_t num_leaves) const {
if (split.loss_chg <= kRtEps) return false;
auto is_zero = [](auto const& sum) {

2
src/tree/param.h
View File

@ -401,7 +401,7 @@ struct SplitEntryContainer {
/*! \brief split index */
bst_feature_t sindex{0};
bst_float split_value{0.0f};
std::vector<uint32_t> cat_bits;
std::vector<std::uint32_t> cat_bits;
bool is_cat{false};
GradientT left_sum;

7
src/tree/updater_gpu_common.cuh
View File

@ -14,9 +14,7 @@
#include "gpu_hist/histogram.cuh"
#include "param.h"
namespace xgboost {
namespace tree {
namespace xgboost::tree {
struct GPUTrainingParam {
// minimum amount of hessian(weight) allowed in a child
float min_child_weight;
@ -136,5 +134,4 @@ struct SumCallbackOp {
return old_prefix;
}
};
} // namespace tree
} // namespace xgboost
} // namespace xgboost::tree

4
tests/cpp/collective/test_comm.cc
View File

@ -4,7 +4,9 @@
#include <gtest/gtest.h>
#include "../../../src/collective/comm.h"
#include "test_worker.h"
#include "../../../src/common/type.h" // for EraseType
#include "test_worker.h" // for TrackerTest
namespace xgboost::collective {
namespace {
class CommTest : public TrackerTest {};

28
tests/cpp/tree/gpu_hist/test_expand_entry.cu Normal file
View File

@ -0,0 +1,28 @@
/**
* Copyright 2023, XGBoost Contributors
*/
#include <gtest/gtest.h>
#include <xgboost/json.h>
#include <xgboost/tree_model.h> // for RegTree
#include "../../../../src/tree/gpu_hist/expand_entry.cuh"
namespace xgboost::tree {
TEST(ExpandEntry, IOGPU) {
DeviceSplitCandidate split;
GPUExpandEntry entry{RegTree::kRoot, 0, split, 3.0, 1.0, 2.0};
Json je{Object{}};
entry.Save(&je);
GPUExpandEntry loaded;
loaded.Load(je);
ASSERT_EQ(entry.base_weight, loaded.base_weight);
ASSERT_EQ(entry.left_weight, loaded.left_weight);
ASSERT_EQ(entry.right_weight, loaded.right_weight);
ASSERT_EQ(entry.GetDepth(), loaded.GetDepth());
ASSERT_EQ(entry.GetLossChange(), loaded.GetLossChange());
}
} // namespace xgboost::tree

57
tests/cpp/tree/hist/test_expand_entry.cc Normal file
View File

@ -0,0 +1,57 @@
/**
* Copyright 2023, XGBoost Contributors
*/
#include <gtest/gtest.h>
#include <xgboost/json.h> // for Json
#include <xgboost/tree_model.h> // for RegTree
#include "../../../../src/tree/hist/expand_entry.h"
namespace xgboost::tree {
TEST(ExpandEntry, IO) {
CPUExpandEntry entry{RegTree::kRoot, 0};
entry.split.Update(1.0, 1, /*new_split_value=*/0.3, true, true, GradStats{1.0, 1.0},
GradStats{2.0, 2.0});
bst_bin_t n_bins_feature = 256;
auto n = common::CatBitField::ComputeStorageSize(n_bins_feature);
entry.split.cat_bits = decltype(entry.split.cat_bits)(n, 0);
common::CatBitField cat_bits{entry.split.cat_bits};
cat_bits.Set(n_bins_feature / 2);
Json je{Object{}};
entry.Save(&je);
CPUExpandEntry loaded;
loaded.Load(je);
ASSERT_EQ(loaded.split.is_cat, entry.split.is_cat);
ASSERT_EQ(loaded.split.cat_bits, entry.split.cat_bits);
ASSERT_EQ(loaded.split.left_sum.GetGrad(), entry.split.left_sum.GetGrad());
ASSERT_EQ(loaded.split.right_sum.GetHess(), entry.split.right_sum.GetHess());
}
TEST(ExpandEntry, IOMulti) {
MultiExpandEntry entry{RegTree::kRoot, 0};
auto left_sum = std::vector<GradientPairPrecise>{{1.0, 1.0}, {1.0, 1.0}};
auto right_sum = std::vector<GradientPairPrecise>{{2.0, 2.0}, {2.0, 2.0}};
entry.split.Update(1.0, 1, /*new_split_value=*/0.3, true, true,
linalg::MakeVec(left_sum.data(), left_sum.size()),
linalg::MakeVec(right_sum.data(), right_sum.size()));
bst_bin_t n_bins_feature = 256;
auto n = common::CatBitField::ComputeStorageSize(n_bins_feature);
entry.split.cat_bits = decltype(entry.split.cat_bits)(n, 0);
common::CatBitField cat_bits{entry.split.cat_bits};
cat_bits.Set(n_bins_feature / 2);
Json je{Object{}};
entry.Save(&je);
MultiExpandEntry loaded;
loaded.Load(je);
ASSERT_EQ(loaded.split.is_cat, entry.split.is_cat);
ASSERT_EQ(loaded.split.cat_bits, entry.split.cat_bits);
ASSERT_EQ(loaded.split.left_sum, entry.split.left_sum);
ASSERT_EQ(loaded.split.right_sum, entry.split.right_sum);
}
} // namespace xgboost::tree