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

De-duplicate GPU parameters. (#4454)

Browse Source 
* Only define `gpu_id` and `n_gpus` in `LearnerTrainParam`
* Pass LearnerTrainParam through XGBoost vid factory method.
* Disable all GPU usage when GPU related parameters are not specified (fixes XGBoost choosing GPU over aggressively).
* Test learner train param io.
* Fix gpu pickling.
This commit is contained in:
Jiaming Yuan
2019-05-29 11:55:57 +08:00
committed by GitHub
parent a3fedbeaa8
commit c589eff941
69 changed files with 927 additions and 562 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
src/common/common.cc
View File

@@ -1,9 +1,10 @@
/*!
* Copyright 2015-2018 by Contributors
* Copyright 2015-2019 by Contributors
* \file common.cc
* \brief Enable all kinds of global variables in common.
*/
#include <dmlc/thread_local.h>
#include <xgboost/logging.h>
#include "common.h"
#include "./random.h"
@@ -29,4 +30,39 @@ int AllVisibleImpl::AllVisible() {
}
#endif // !defined(XGBOOST_USE_CUDA)
constexpr GPUSet::GpuIdType GPUSet::kAll;
GPUSet GPUSet::All(GpuIdType gpu_id, GpuIdType n_gpus, int32_t n_rows) {
CHECK_GE(gpu_id, 0) << "gpu_id must be >= 0.";
CHECK_GE(n_gpus, -1) << "n_gpus must be >= -1.";
GpuIdType const n_devices_visible = AllVisible().Size();
if (n_devices_visible == 0 || n_gpus == 0 || n_rows == 0) {
LOG(DEBUG) << "Runing on CPU.";
return Empty();
}
GpuIdType const n_available_devices = n_devices_visible - gpu_id;
if (n_gpus == kAll) { // Use all devices starting from `gpu_id'.
CHECK(gpu_id < n_devices_visible)
<< "\ngpu_id should be less than number of visible devices.\ngpu_id: "
<< gpu_id
<< ", number of visible devices: "
<< n_devices_visible;
GpuIdType n_devices =
n_available_devices < n_rows ? n_available_devices : n_rows;
LOG(DEBUG) << "GPU ID: " << gpu_id << ", Number of GPUs: " << n_devices;
return Range(gpu_id, n_devices);
} else { // Use devices in ( gpu_id, gpu_id + n_gpus ).
CHECK_LE(n_gpus, n_available_devices)
<< "Starting from gpu id: " << gpu_id << ", there are only "
<< n_available_devices << " available devices, while n_gpus is set to: "
<< n_gpus;
GpuIdType n_devices = n_gpus < n_rows ? n_gpus : n_rows;
LOG(DEBUG) << "GPU ID: " << gpu_id << ", Number of GPUs: " << n_devices;
return Range(gpu_id, n_devices);
}
}
} // namespace xgboost

30
src/common/common.h
View File

@@ -135,8 +135,8 @@ class Range {
Iterator begin_;
Iterator end_;
};
} // namespace common
struct AllVisibleImpl {
static int AllVisible();
};
@@ -160,33 +160,7 @@ class GPUSet {
}
/*! \brief n_gpus and num_rows both are upper bounds. */
static GPUSet All(GpuIdType gpu_id, GpuIdType n_gpus,
GpuIdType num_rows = std::numeric_limits<GpuIdType>::max()) {
CHECK_GE(gpu_id, 0) << "gpu_id must be >= 0.";
CHECK_GE(n_gpus, -1) << "n_gpus must be >= -1.";
GpuIdType const n_devices_visible = AllVisible().Size();
if (n_devices_visible == 0 || n_gpus == 0) { return Empty(); }
GpuIdType const n_available_devices = n_devices_visible - gpu_id;
if (n_gpus == kAll) { // Use all devices starting from `gpu_id'.
CHECK(gpu_id < n_devices_visible)
<< "\ngpu_id should be less than number of visible devices.\ngpu_id: "
<< gpu_id
<< ", number of visible devices: "
<< n_devices_visible;
GpuIdType n_devices =
n_available_devices < num_rows ? n_available_devices : num_rows;
return Range(gpu_id, n_devices);
} else { // Use devices in ( gpu_id, gpu_id + n_gpus ).
CHECK_LE(n_gpus, n_available_devices)
<< "Starting from gpu id: " << gpu_id << ", there are only "
<< n_available_devices << " available devices, while n_gpus is set to: "
<< n_gpus;
GpuIdType n_devices = n_gpus < num_rows ? n_gpus : num_rows;
return Range(gpu_id, n_devices);
}
}
GpuIdType num_rows = std::numeric_limits<GpuIdType>::max());
static GPUSet AllVisible() {
GpuIdType n = AllVisibleImpl::AllVisible();

81
src/common/enum_class_param.h
View File

@@ -1,81 +0,0 @@
/*!
* Copyright 2018 by Contributors
* \file enum_class_param.h
* \brief macro for using C++11 enum class as DMLC parameter
* \author Hyunsu Philip Cho
*/
#ifndef XGBOOST_COMMON_ENUM_CLASS_PARAM_H_
#define XGBOOST_COMMON_ENUM_CLASS_PARAM_H_
#include <dmlc/parameter.h>
#include <string>
#include <type_traits>
/*!
* \brief Specialization of FieldEntry for enum class (backed by int)
*
* Use this macro to use C++11 enum class as DMLC parameters
*
* Usage:
*
* \code{.cpp}
*
* // enum class must inherit from int type
* enum class Foo : int {
* kBar = 0, kFrog = 1, kCat = 2, kDog = 3
* };
*
* // This line is needed to prevent compilation error
* DECLARE_FIELD_ENUM_CLASS(Foo);
*
* // Now define DMLC parameter as usual;
* // enum classes can now be members.
* struct MyParam : dmlc::Parameter<MyParam> {
* Foo foo;
* DMLC_DECLARE_PARAMETER(MyParam) {
* DMLC_DECLARE_FIELD(foo)
* .set_default(Foo::kBar)
* .add_enum("bar", Foo::kBar)
* .add_enum("frog", Foo::kFrog)
* .add_enum("cat", Foo::kCat)
* .add_enum("dog", Foo::kDog);
* }
* };
*
* DMLC_REGISTER_PARAMETER(MyParam);
* \endcode
*/
#define DECLARE_FIELD_ENUM_CLASS(EnumClass) \
namespace dmlc { \
namespace parameter { \
template <> \
class FieldEntry<EnumClass> : public FieldEntry<int> { \
public: \
FieldEntry<EnumClass>() { \
static_assert( \
std::is_same<int, typename std::underlying_type<EnumClass>::type>::value, \
"enum class must be backed by int"); \
is_enum_ = true; \
} \
using Super = FieldEntry<int>; \
void Set(void *head, const std::string &value) const override { \
Super::Set(head, value); \
} \
inline FieldEntry<EnumClass>& add_enum(const std::string &key, EnumClass value) { \
Super::add_enum(key, static_cast<int>(value)); \
return *this; \
} \
inline FieldEntry<EnumClass>& set_default(const EnumClass& default_value) { \
default_value_ = static_cast<int>(default_value); \
has_default_ = true; \
return *this; \
} \
inline void Init(const std::string &key, void *head, EnumClass& ref) { /* NOLINT */ \
Super::Init(key, head, *reinterpret_cast<int*>(&ref)); \
} \
}; \
} /* namespace parameter */ \
} /* namespace dmlc */
#endif // XGBOOST_COMMON_ENUM_CLASS_PARAM_H_

9
src/common/hist_util.cu
View File

@@ -383,8 +383,8 @@ struct GPUSketcher {
hmat->Init(&sketches, param_.max_bin);
}
GPUSketcher(tree::TrainParam param, size_t n_rows) : param_(std::move(param)) {
dist_ = GPUDistribution::Block(GPUSet::All(param_.gpu_id, param_.n_gpus, n_rows));
GPUSketcher(tree::TrainParam param, GPUSet const& devices) : param_(std::move(param)) {
dist_ = GPUDistribution::Block(devices);
}
private:
@@ -395,8 +395,9 @@ struct GPUSketcher {
void DeviceSketch
(const SparsePage& batch, const MetaInfo& info,
const tree::TrainParam& param, HistCutMatrix* hmat, int gpu_batch_nrows) {
GPUSketcher sketcher(param, info.num_row_);
const tree::TrainParam& param, HistCutMatrix* hmat, int gpu_batch_nrows,
GPUSet const& devices) {
GPUSketcher sketcher(param, devices);
sketcher.Sketch(batch, info, hmat, gpu_batch_nrows);
}

3
src/common/hist_util.h
View File

@@ -118,7 +118,8 @@ struct HistCutMatrix {
/*! \brief Builds the cut matrix on the GPU */
void DeviceSketch
(const SparsePage& batch, const MetaInfo& info,
const tree::TrainParam& param, HistCutMatrix* hmat, int gpu_batch_nrows);
const tree::TrainParam& param, HistCutMatrix* hmat, int gpu_batch_nrows,
GPUSet const& devices);
/*!
* \brief A single row in global histogram index.

2
src/common/host_device_vector.h
View File

@@ -1,5 +1,5 @@
/*!
* Copyright 2017 XGBoost contributors
* Copyright 2017-2019 XGBoost contributors
*/
/**