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

Complete cudf support. (#4850)

Browse Source 
* Handles missing value.
* Accept all floating point and integer types.
* Move to cudf 9.0 API.
* Remove requirement on `null_count`.
* Arbitrary column types support.
This commit is contained in:
Jiaming Yuan
2019-09-16 23:52:00 -04:00
committed by GitHub
parent 125bcec62e
commit 5374f52531
17 changed files with 702 additions and 339 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/c_api/c_api.cc
View File

@@ -190,11 +190,11 @@ int XGDMatrixCreateFromDataIter(
}
XGB_DLL int XGDMatrixCreateFromArrayInterfaces(
char const* c_json_strs, DMatrixHandle* out) {
char const* c_json_strs, bst_int has_missing, bst_float missing, DMatrixHandle* out) {
API_BEGIN();
std::string json_str {c_json_strs};
std::unique_ptr<data::SimpleCSRSource> source (new data::SimpleCSRSource());
source->CopyFrom(json_str);
source->CopyFrom(json_str, has_missing, missing);
*out = new std::shared_ptr<DMatrix>(DMatrix::Create(std::move(source)));
API_END();
}

2
src/common/device_helpers.cuh
View File

@@ -345,6 +345,8 @@ struct XGBCachingDeviceAllocatorImpl : thrust::device_malloc_allocator<T> {
// Replacement of allocator with custom backend should occur here
template <typename T>
using XGBDeviceAllocator = detail::XGBDefaultDeviceAllocatorImpl<T>;
/*! Be careful that the initialization constructor is a no-op, which means calling
* `vec.resize(n, 1)` won't initialize the memory region to 1. */
template <typename T>
using XGBCachingDeviceAllocator = detail::XGBCachingDeviceAllocatorImpl<T>;
/** \brief Specialisation of thrust device vector using custom allocator. */

57
src/common/math.h
View File

@@ -9,11 +9,11 @@
#include <xgboost/base.h>
#include <algorithm>
#include <cmath>
#include <limits>
#include <utility>
#include <vector>
#include <cmath>
#include <algorithm>
#include <utility>
namespace xgboost {
namespace common {
@@ -26,6 +26,23 @@ XGBOOST_DEVICE inline float Sigmoid(float x) {
return 1.0f / (1.0f + expf(-x));
}
/*!
* \brief Equality test for both integer and floating point.
*/
template <typename T, typename U>
XGBOOST_DEVICE constexpr bool CloseTo(T a, U b) {
using Casted =
typename std::conditional<
std::is_floating_point<T>::value || std::is_floating_point<U>::value,
double,
typename std::conditional<
std::is_signed<T>::value || std::is_signed<U>::value,
int64_t,
uint64_t>::type>::type;
return std::is_floating_point<Casted>::value ?
std::abs(static_cast<Casted>(a) -static_cast<Casted>(b)) < 1e-6 : a == b;
}
/*!
* \brief Do inplace softmax transformaton on start to end
*
@@ -119,14 +136,34 @@ inline static bool CmpSecond(const std::pair<float, unsigned> &a,
// check nan
bool CheckNAN(double v);
#else
template<typename T>
inline bool CheckNAN(T v) {
#ifdef _MSC_VER
return (_isnan(v) != 0);
#else
return std::isnan(v);
#endif // _MSC_VER
// Redefined here to workaround a VC bug that doesn't support overloadng for integer
// types.
template <typename T>
XGBOOST_DEVICE typename std::enable_if<
std::numeric_limits<T>::is_integer, bool>::type
CheckNAN(T) {
return false;
}
XGBOOST_DEVICE bool inline CheckNAN(float x) {
#if (defined(_WIN32) || defined(__CUDA_ARCH__)) && \
!defined(__MINGW64__) && !defined(__MINGW32__) && !defined(__CYGWIN__)
return isnan(x);
#else
return std::isnan(x);
#endif // has c++11 std::isnan
}
XGBOOST_DEVICE bool inline CheckNAN(double x) {
#if (defined(_WIN32) || defined(__CUDA_ARCH__)) && \
!defined(__MINGW64__) && !defined(__MINGW32__) && !defined(__CYGWIN__)
return isnan(x);
#else
return std::isnan(x);
#endif // has c++11 std::isnan
}
#endif // XGBOOST_STRICT_R_MODE_
// GPU version is not uploaded in CRAN anyway.

192
src/data/columnar.h
View File

@@ -12,18 +12,20 @@
#include "xgboost/data.h"
#include "xgboost/json.h"
#include "xgboost/logging.h"
#include "../common/span.h"
#include "../common/bitfield.h"
namespace xgboost {
// A view over __array_interface__
template <typename T>
struct Columnar {
using mask_type = unsigned char;
using index_type = int32_t;
common::Span<float> data;
common::Span<T> data;
RBitField8 valid;
int32_t size;
int32_t null_count;
};
// Common errors in parsing columnar format.
@@ -49,13 +51,7 @@ struct ColumnarErrors {
static char const* Version() {
return "Only version 1 of __cuda_array_interface__ is being supported.";
}
static char const* toFloat() {
return "Please convert the input into float32 first.";
}
static char const* toUInt() {
return "Please convert the Group into unsigned 32 bit integers first.";
}
static char const* ofType(std::string type) {
static char const* ofType(std::string const& type) {
static std::string str;
str.clear();
str += " should be of ";
@@ -63,20 +59,176 @@ struct ColumnarErrors {
str += " type.";
return str.c_str();
}
static std::string UnknownTypeStr(std::string const& typestr) {
return "typestr from array interface: " + typestr + " is not supported.";
}
};
template <typename PtrType>
PtrType GetPtrFromArrayData(std::map<std::string, Json> const& obj) {
if (obj.find("data") == obj.cend()) {
LOG(FATAL) << "Empty data passed in.";
// TODO(trivialfis): Abstract this into a class that accept a json
// object and turn it into an array (for cupy and numba).
class ArrayInterfaceHandler {
public:
template <typename T>
static constexpr char TypeChar() {
return
(std::is_floating_point<T>::value ? 'f' :
(std::is_integral<T>::value ?
(std::is_signed<T>::value ? 'i' : 'u') : '\0'));
}
static std::string TypeStr(char c) {
switch (c) {
case 't':
return "Bit field";
case 'b':
return "Boolean";
case 'i':
return "Integer";
case 'u':
return "Unsigned integer";
case 'f':
return "Floating point";
default:
LOG(FATAL) << "Invalid type code: " << c << " in typestr of input array interface.";
return "";
}
}
template <typename PtrType>
static PtrType GetPtrFromArrayData(std::map<std::string, Json> const& obj) {
if (obj.find("data") == obj.cend()) {
LOG(FATAL) << "Empty data passed in.";
}
auto p_data = reinterpret_cast<PtrType>(static_cast<size_t>(
get<Integer const>(
get<Array const>(
obj.at("data"))
.at(0))));
return p_data;
}
static void Validate(std::map<std::string, Json> const& array) {
if (array.find("version") == array.cend()) {
LOG(FATAL) << "Missing version field for array interface";
}
auto version = get<Integer const>(array.at("version"));
CHECK_EQ(version, 1) << ColumnarErrors::Version();
if (array.find("typestr") == array.cend()) {
LOG(FATAL) << "Missing typestr field for array interface";
}
auto typestr = get<String const>(array.at("typestr"));
CHECK_EQ(typestr.size(), 3) << ColumnarErrors::TypestrFormat();
CHECK_NE(typestr.front(), '>') << ColumnarErrors::BigEndian();
if (array.find("shape") == array.cend()) {
LOG(FATAL) << "Missing shape field for array interface";
}
if (array.find("data") == array.cend()) {
LOG(FATAL) << "Missing data field for array interface";
}
}
// Find null mask (validity mask) field
// Mask object is also an array interface, but with different requirements.
static void ExtractMask(std::map<std::string, Json> const& column,
common::Span<RBitField8::value_type>* p_out) {
auto& s_mask = *p_out;
if (column.find("mask") != column.cend()) {
auto const& j_mask = get<Object const>(column.at("mask"));
Validate(j_mask);
auto p_mask = GetPtrFromArrayData<RBitField8::value_type*>(j_mask);
auto j_shape = get<Array const>(j_mask.at("shape"));
CHECK_EQ(j_shape.size(), 1) << ColumnarErrors::Dimension(1);
CHECK_EQ(get<Integer>(j_shape.front()) % 8, 0) <<
"Length of validity mask must be a multiple of 8 bytes.";
int64_t size = get<Integer>(j_shape.at(0)) *
sizeof(unsigned char) / sizeof(RBitField8::value_type);
auto typestr = get<String const>(j_mask.at("typestr"));
if (typestr.at(1) == 't') {
CHECK_EQ(typestr.at(2), '1') << "There can be only 1 bit in each entry of bitfield.";
} else if (typestr.at(1) == 'i') {
CHECK_EQ(typestr.at(2), '1') << "mask with integer type should be of 1 byte per integer.";
} else {
LOG(FATAL) << "mask must be of integer type or bit field type.";
}
// For now this is just 1
int64_t const type_length = typestr.at(2) - 48;
s_mask = {p_mask, size / type_length};
}
}
template <typename T>
static common::Span<T> ExtractData(std::map<std::string, Json> const& column) {
Validate(column);
auto typestr = get<String const>(column.at("typestr"));
CHECK_EQ(typestr.at(1), TypeChar<T>())
<< "Input data type and typestr mismatch. typestr: " << typestr;
CHECK_EQ(typestr.at(2), static_cast<char>(sizeof(T) + 48))
<< "Input data type and typestr mismatch. typestr: " << typestr;
auto j_shape = get<Array const>(column.at("shape"));
CHECK_EQ(j_shape.size(), 1) << ColumnarErrors::Dimension(1);
if (column.find("strides") != column.cend()) {
auto strides = get<Array const>(column.at("strides"));
CHECK_EQ(strides.size(), 1) << ColumnarErrors::Dimension(1);
CHECK_EQ(get<Integer>(strides.at(0)), 4) << ColumnarErrors::Contigious();
}
auto length = get<Integer const>(j_shape.at(0));
T* p_data = ArrayInterfaceHandler::GetPtrFromArrayData<T*>(column);
return common::Span<T>{p_data, length};
}
template <typename T>
static Columnar<T> ExtractArray(std::map<std::string, Json> const& column) {
common::Span<T> s_data { ArrayInterfaceHandler::ExtractData<T>(column) };
Columnar<T> foreign_col;
foreign_col.data = s_data;
foreign_col.size = s_data.size();
common::Span<RBitField8::value_type> s_mask;
ArrayInterfaceHandler::ExtractMask(column, &s_mask);
foreign_col.valid = RBitField8(s_mask);
return foreign_col;
}
};
#define DISPATCH_TYPE(__dispatched_func, __typestr, ...) { \
if (__typestr.at(1) == 'f' && __typestr.at(2) == '4') { \
__dispatched_func<float>(__VA_ARGS__); \
} else if (__typestr.at(1) == 'f' && __typestr.at(2) == '8') { \
__dispatched_func<double>(__VA_ARGS__); \
} else if (__typestr.at(1) == 'i' && __typestr.at(2) == '1') { \
__dispatched_func<int8_t>(__VA_ARGS__); \
} else if (__typestr.at(1) == 'i' && __typestr.at(2) == '2') { \
__dispatched_func<int16_t>(__VA_ARGS__); \
} else if (__typestr.at(1) == 'i' && __typestr.at(2) == '4') { \
__dispatched_func<int32_t>(__VA_ARGS__); \
} else if (__typestr.at(1) == 'i' && __typestr.at(2) == '8') { \
__dispatched_func<int64_t>(__VA_ARGS__); \
} else if (__typestr.at(1) == 'u' && __typestr.at(2) == '1') { \
__dispatched_func<uint8_t>(__VA_ARGS__); \
} else if (__typestr.at(1) == 'u' && __typestr.at(2) == '2') { \
__dispatched_func<uint16_t>(__VA_ARGS__); \
} else if (__typestr.at(1) == 'u' && __typestr.at(2) == '4') { \
__dispatched_func<uint32_t>(__VA_ARGS__); \
} else if (__typestr.at(1) == 'u' && __typestr.at(2) == '8') { \
__dispatched_func<uint64_t>(__VA_ARGS__); \
} else { \
LOG(FATAL) << ColumnarErrors::UnknownTypeStr(__typestr); \
} \
}
auto p_data = reinterpret_cast<PtrType>(static_cast<size_t>(
get<Integer const>(
get<Array const>(
obj.at("data"))
.at(0))));
return p_data;
}
} // namespace xgboost
#endif // XGBOOST_DATA_COLUMNAR_H_

85
src/data/data.cu
View File

@@ -2,8 +2,8 @@
* Copyright 2019 by XGBoost Contributors
*
* \file data.cu
* \brief Handles setting metainfo from array interface.
*/
#include "xgboost/data.h"
#include "xgboost/logging.h"
#include "xgboost/json.h"
@@ -12,75 +12,60 @@
namespace xgboost {
template <typename T>
void CopyInfoImpl(std::map<std::string, Json> const& column, HostDeviceVector<float>* out) {
auto SetDeviceToPtr = [](void* ptr) {
cudaPointerAttributes attr;
dh::safe_cuda(cudaPointerGetAttributes(&attr, ptr));
int32_t ptr_device = attr.device;
dh::safe_cuda(cudaSetDevice(ptr_device));
return ptr_device;
};
common::Span<T> s_data { ArrayInterfaceHandler::ExtractData<T>(column) };
auto ptr_device = SetDeviceToPtr(s_data.data());
thrust::device_ptr<T> p_src {s_data.data()};
auto length = s_data.size();
out->SetDevice(ptr_device);
out->Resize(length);
auto p_dst = thrust::device_pointer_cast(out->DevicePointer());
thrust::copy(p_src, p_src + length, p_dst);
}
void MetaInfo::SetInfo(const char * c_key, std::string const& interface_str) {
Json j_arr = Json::Load({interface_str.c_str(), interface_str.size()});
auto const& j_arr_obj = get<Object>(j_arr);
std::string key {c_key};
auto version = get<Integer const>(j_arr_obj.at("version"));
CHECK_EQ(version, 1) << ColumnarErrors::Version();
ArrayInterfaceHandler::Validate(j_arr_obj);
if (j_arr_obj.find("mask") != j_arr_obj.cend()) {
LOG(FATAL) << "Meta info " << key << " should be dense, found validity mask";
}
auto const& typestr = get<String const>(j_arr_obj.at("typestr"));
auto typestr = get<String const>(j_arr_obj.at("typestr"));
CHECK_EQ(typestr.size(), 3) << ColumnarErrors::TypestrFormat();
CHECK_NE(typestr.front(), '>') << ColumnarErrors::BigEndian();
auto j_shape = get<Array const>(j_arr_obj.at("shape"));
CHECK_EQ(j_shape.size(), 1) << ColumnarErrors::Dimension(1);
auto length = get<Integer const>(j_shape.at(0));
CHECK_GT(length, 0) << "Label set cannot be empty.";
if (j_arr_obj.find("strides") != j_arr_obj.cend()) {
auto strides = get<Array const>(j_arr_obj.at("strides"));
CHECK_EQ(get<Integer>(strides.at(0)), 4) << ColumnarErrors::Contigious();
}
float* p_data = GetPtrFromArrayData<float*>(j_arr_obj);
cudaPointerAttributes attr;
dh::safe_cuda(cudaPointerGetAttributes(&attr, p_data));
int32_t ptr_device = attr.device;
dh::safe_cuda(cudaSetDevice(ptr_device));
thrust::device_ptr<float> p_src {p_data};
HostDeviceVector<float>* dst;
if (key == "root_index") {
LOG(FATAL) << "root index for columnar data is not supported.";
} else if (key == "label") {
dst = &labels_;
CHECK_EQ(typestr.at(1), 'f') << "Label"
<< ColumnarErrors::ofType("floating point");
CHECK_EQ(typestr.at(2), '4') << ColumnarErrors::toFloat();
DISPATCH_TYPE(CopyInfoImpl, typestr, j_arr_obj, &labels_);
} else if (key == "weight") {
dst = &weights_;
CHECK_EQ(typestr.at(1), 'f') << "Weight"
<< ColumnarErrors::ofType("floating point");;
CHECK_EQ(typestr.at(2), '4') << ColumnarErrors::toFloat();
DISPATCH_TYPE(CopyInfoImpl, typestr, j_arr_obj, &weights_);
} else if (key == "base_margin") {
dst = &base_margin_;
CHECK_EQ(typestr.at(1), 'f') << "Base Margin"
<< ColumnarErrors::ofType("floating point");
CHECK_EQ(typestr.at(2), '4') << ColumnarErrors::toFloat();
DISPATCH_TYPE(CopyInfoImpl, typestr, j_arr_obj, &base_margin_);
} else if (key == "group") {
CHECK_EQ(typestr.at(1), 'u') << "Group"
<< ColumnarErrors::ofType("unsigned 32 bit integers");
CHECK_EQ(typestr.at(2), '4') << ColumnarErrors::toUInt();
// Ranking is not performed on device.
auto s_data = ArrayInterfaceHandler::ExtractData<uint32_t>(j_arr_obj);
thrust::device_ptr<uint32_t> p_src {s_data.data()};
auto length = s_data.size();
group_ptr_.resize(length + 1);
group_ptr_[0] = 0;
// Ranking is not performed on device.
thrust::copy(p_src, p_src + length, group_ptr_.begin() + 1);
for (size_t i = 1; i < group_ptr_.size(); ++i) {
group_ptr_[i] = group_ptr_[i - 1] + group_ptr_[i];
}
std::partial_sum(group_ptr_.begin(), group_ptr_.end(), group_ptr_.begin());
return;
} else {
LOG(FATAL) << "Unknown metainfo: " << key;
}
dst->SetDevice(ptr_device);
dst->Resize(length);
auto p_dst = thrust::device_pointer_cast(dst->DevicePointer());
thrust::copy(p_src, p_src + length, p_dst);
}
} // namespace xgboost

113
src/data/simple_csr_source.cc
View File

@@ -123,15 +123,29 @@ const SparsePage& SimpleCSRSource::Value() const {
/*!
* Please be careful that, in official specification, the only three required fields are
* `shape', `version' and `typestr'. Any other is optional, including `data'. But here
* we have two additional requirements for input data:
* we have one additional requirements for input data:
*
* - `data' field is required, passing in an empty dataset is not accepted, as most (if
* not all) of our algorithms don't have test for empty dataset. An error is better
* than a crash.
*
* - `null_count' is required when `mask' is presented. We can compute `null_count'
* ourselves and copy the result back to host for memory allocation. But it's in the
* specification of Apache Arrow hence it should be readily available,
* Missing value handling:
* Missing value is specified:
* - Ignore the validity mask from columnar format.
* - Remove entries that equals to missing value.
* - missing = NaN:
* - Remove entries that is NaN
* - missing != NaN:
* - Check for NaN entries, throw an error if found.
* Missing value is not specified:
* - Remove entries that is specifed as by validity mask.
* - Remove NaN entries.
*
* What if invalid value from dataframe is 0 but I specify missing=NaN in XGBoost? Since
* validity mask is ignored, all 0s are preserved in XGBoost.
*
* FIXME(trivialfis): Put above into document after we have a consistent way for
* processing input data.
*
* Sample input:
* [
@@ -160,100 +174,29 @@ const SparsePage& SimpleCSRSource::Value() const {
* false
* ],
* "typestr": "|i1",
* "version": 1,
* "null_count": 1
* "version": 1
* }
* }
* ]
*/
void SimpleCSRSource::CopyFrom(std::string const& cuda_interfaces_str) {
void SimpleCSRSource::CopyFrom(std::string const& cuda_interfaces_str,
bool has_missing, float missing) {
Json interfaces = Json::Load({cuda_interfaces_str.c_str(),
cuda_interfaces_str.size()});
std::vector<Json> const& columns = get<Array>(interfaces);
size_t n_columns = columns.size();
CHECK_GT(n_columns, 0);
CHECK_GT(n_columns, 0) << "Number of columns must not be greater than 0.";
std::vector<Columnar> foreign_cols(n_columns);
for (size_t i = 0; i < columns.size(); ++i) {
CHECK(IsA<Object>(columns[i]));
auto const& column = get<Object const>(columns[i]);
auto const& typestr = get<String const>(columns[0]["typestr"]);
CHECK_EQ(typestr.size(), 3) << ColumnarErrors::TypestrFormat();
CHECK_NE(typestr.front(), '>') << ColumnarErrors::BigEndian();
auto version = get<Integer const>(column.at("version"));
CHECK_EQ(version, 1) << ColumnarErrors::Version();
// Find null mask (validity mask) field
// Mask object is also an array interface, but with different requirements.
// TODO(trivialfis): Abstract this into a class that accept a json
// object and turn it into an array (for cupy and numba).
common::Span<RBitField8::value_type> s_mask;
int32_t null_count {0};
if (column.find("mask") != column.cend()) {
auto const& j_mask = get<Object const>(column.at("mask"));
auto p_mask = GetPtrFromArrayData<RBitField8::value_type*>(j_mask);
auto j_shape = get<Array const>(j_mask.at("shape"));
CHECK_EQ(j_shape.size(), 1) << ColumnarErrors::Dimension(1);
CHECK_EQ(get<Integer>(j_shape.front()) % 8, 0) <<
"Length of validity map must be a multiple of 8 bytes.";
int64_t size = get<Integer>(j_shape.at(0)) *
sizeof(unsigned char) / sizeof(RBitField8::value_type);
s_mask = {p_mask, size};
auto typestr = get<String const>(j_mask.at("typestr"));
CHECK_EQ(typestr.size(), 3) << ColumnarErrors::TypestrFormat();
CHECK_NE(typestr.front(), '>') << ColumnarErrors::BigEndian();
CHECK_EQ(typestr.at(1), 'i') << "mask" << ColumnarErrors::ofType("unsigned char");
CHECK_EQ(typestr.at(2), '1') << "mask" << ColumnarErrors::toUInt();
CHECK(j_mask.find("null_count") != j_mask.cend()) <<
"Column with null mask must include null_count as "
"part of mask object for XGBoost.";
null_count = get<Integer const>(j_mask.at("null_count"));
}
// Find data field
if (column.find("data") == column.cend()) {
LOG(FATAL) << "Empty dataset passed in.";
}
auto typestr = get<String const>(column.at("typestr"));
CHECK_EQ(typestr.size(), 3) << ColumnarErrors::TypestrFormat();
CHECK_NE(typestr.front(), '>') << ColumnarErrors::BigEndian();
CHECK_EQ(typestr.at(1), 'f') << "data" << ColumnarErrors::ofType("floating point");
CHECK_EQ(typestr.at(2), '4') << ColumnarErrors::toFloat();
auto j_shape = get<Array const>(column.at("shape"));
CHECK_EQ(j_shape.size(), 1) << ColumnarErrors::Dimension(1);
if (column.find("strides") != column.cend()) {
auto strides = get<Array const>(column.at("strides"));
CHECK_EQ(strides.size(), 1) << ColumnarErrors::Dimension(1);
CHECK_EQ(get<Integer>(strides.at(0)), 4) << ColumnarErrors::Contigious();
}
auto length = get<Integer const>(j_shape.at(0));
float* p_data = GetPtrFromArrayData<float*>(column);
common::Span<float> s_data {p_data, length};
foreign_cols[i].data = s_data;
foreign_cols[i].valid = RBitField8(s_mask);
foreign_cols[i].size = s_data.size();
foreign_cols[i].null_count = null_count;
}
info.num_col_ = n_columns;
info.num_row_ = foreign_cols[0].size;
for (size_t i = 0; i < n_columns; ++i) {
CHECK_EQ(foreign_cols[0].size, foreign_cols[i].size);
info.num_nonzero_ += foreign_cols[i].data.size() - foreign_cols[i].null_count;
}
this->FromDeviceColumnar(foreign_cols);
this->FromDeviceColumnar(columns, has_missing, missing);
}
#if !defined(XGBOOST_USE_CUDA)
void SimpleCSRSource::FromDeviceColumnar(std::vector<Columnar> cols) {
void SimpleCSRSource::FromDeviceColumnar(std::vector<Json> const& columns,
bool has_missing, float missing) {
LOG(FATAL) << "XGBoost version is not compiled with GPU support";
}
#endif // !defined(XGBOOST_USE_CUDA)

198
src/data/simple_csr_source.cu
View File

@@ -12,102 +12,190 @@
#include <xgboost/base.h>
#include <xgboost/data.h>
#include <cmath>
#include <vector>
#include <algorithm>
#include "simple_csr_source.h"
#include "columnar.h"
#include "../common/math.h"
#include "../common/bitfield.h"
#include "../common/device_helpers.cuh"
namespace xgboost {
namespace data {
template <size_t kBlockThreads>
__global__ void CountValidKernel(common::Span<Columnar const> columns,
int32_t const n_rows,
common::Span<size_t> offsets) {
// One block for a column
auto const bid = blockIdx.x;
auto const tid = threadIdx.x;
if (bid >= columns.size()) {
template <typename T>
__global__ void CountValidKernel(Columnar<T> const column,
bool has_missing, float missing,
int32_t* flag, common::Span<size_t> offsets) {
auto const tid = threadIdx.x + blockDim.x * blockIdx.x;
bool const missing_is_nan = common::CheckNAN(missing);
if (tid >= column.size) {
return;
}
RBitField8 const mask = columns[bid].valid;
for (auto r = tid; r < n_rows; r += kBlockThreads) {
if (mask.Data() == nullptr || mask.Check(r)) {
atomicAdd(reinterpret_cast<BitFieldAtomicType*>(&offsets[r+1]),
static_cast<BitFieldAtomicType>(1));
RBitField8 const mask = column.valid;
if (!has_missing) {
if ((mask.Data() == nullptr || mask.Check(tid)) &&
!common::CheckNAN(column.data[tid])) {
offsets[tid+1] += 1;
}
} else if (missing_is_nan) {
if (!common::CheckNAN(column.data[tid])) {
offsets[tid+1] += 1;
}
} else {
if (!common::CloseTo(column.data[tid], missing)) {
offsets[tid+1] += 1;
}
if (common::CheckNAN(column.data[tid])) {
*flag = 1;
}
}
}
__global__ void CreateCSRKernel(Columnar const column,
int32_t colid,
common::Span<size_t> offsets,
common::Span<Entry> out_data) {
auto tid = threadIdx.x + blockDim.x * blockIdx.x;
template <typename T>
__device__ void AssignValue(T fvalue, int32_t colid,
common::Span<size_t> out_offsets, common::Span<Entry> out_data) {
auto const tid = threadIdx.x + blockDim.x * blockIdx.x;
int32_t oid = out_offsets[tid];
out_data[oid].fvalue = fvalue;
out_data[oid].index = colid;
out_offsets[tid] += 1;
}
template <typename T>
__global__ void CreateCSRKernel(Columnar<T> const column,
int32_t colid, bool has_missing, float missing,
common::Span<size_t> offsets, common::Span<Entry> out_data) {
auto const tid = threadIdx.x + blockDim.x * blockIdx.x;
if (column.size <= tid) {
return;
}
if (column.valid.Data() == nullptr || column.valid.Check(tid)) {
int32_t oid = offsets[tid];
out_data[oid].fvalue = column.data[tid];
out_data[oid].index = colid;
offsets[tid] += 1;
bool const missing_is_nan = common::CheckNAN(missing);
if (!has_missing) {
// no missing value is specified
if ((column.valid.Data() == nullptr || column.valid.Check(tid)) &&
!common::CheckNAN(column.data[tid])) {
AssignValue(column.data[tid], colid, offsets, out_data);
}
} else if (missing_is_nan) {
// specified missing value, but it's NaN
if (!common::CheckNAN(column.data[tid])) {
AssignValue(column.data[tid], colid, offsets, out_data);
}
} else {
// specified missing value, and it's not NaN
if (!common::CloseTo(column.data[tid], missing)) {
AssignValue(column.data[tid], colid, offsets, out_data);
}
}
}
void SimpleCSRSource::FromDeviceColumnar(std::vector<Columnar> cols) {
uint64_t const n_cols = cols.size();
uint64_t const n_rows = cols[0].size;
template <typename T>
void CountValid(std::vector<Json> const& j_columns, uint32_t column_id,
bool has_missing, float missing,
HostDeviceVector<size_t>* out_offset,
dh::caching_device_vector<int32_t>* out_d_flag,
uint32_t* out_n_rows) {
int32_t constexpr kThreads = 256;
auto const& j_column = j_columns[column_id];
auto const& column_obj = get<Object const>(j_column);
Columnar<T> foreign_column = ArrayInterfaceHandler::ExtractArray<T>(column_obj);
uint32_t const n_rows = foreign_column.size;
auto ptr = cols[0].data.data();
auto ptr = foreign_column.data.data();
int32_t device = dh::CudaGetPointerDevice(ptr);
CHECK_NE(device, -1);
for (int32_t i = 1; i < n_cols; ++i) {
auto ptr = cols[i].data.data();
int32_t ptr_device = dh::CudaGetPointerDevice(ptr);
CHECK_EQ(device, ptr_device)
<< "GPU ID at 0^th column: " << device << ", "
<< "GPU ID at column " << i << ": " << ptr_device;
}
dh::safe_cuda(cudaSetDevice(device));
page_.offset.SetDevice(device);
page_.offset.Resize(info.num_row_ + 1);
if (column_id == 0) {
out_offset->SetDevice(device);
out_offset->Resize(n_rows + 1);
}
CHECK_EQ(out_offset->DeviceIdx(), device)
<< "All columns should use the same device.";
CHECK_EQ(out_offset->Size(), n_rows + 1)
<< "All columns should have same number of rows.";
page_.data.SetDevice(device);
page_.data.Resize(info.num_nonzero_);
common::Span<size_t> s_offsets = out_offset->DeviceSpan();
auto s_data = page_.data.DeviceSpan();
auto s_offsets = page_.offset.DeviceSpan();
CHECK_EQ(s_offsets.size(), n_rows + 1);
int32_t const kBlocks = common::DivRoundUp(n_rows, kThreads);
CountValidKernel<T><<<kBlocks, kThreads>>>(
foreign_column,
has_missing, missing,
out_d_flag->data().get(), s_offsets);
*out_n_rows = n_rows;
}
template <typename T>
void CreateCSR(std::vector<Json> const& j_columns, uint32_t column_id, uint32_t n_rows,
bool has_missing, float missing,
dh::device_vector<size_t>* tmp_offset, common::Span<Entry> s_data) {
int32_t constexpr kThreads = 256;
dh::device_vector<Columnar> d_cols(cols);
auto s_d_cols = dh::ToSpan(d_cols);
auto const& j_column = j_columns[column_id];
auto const& column_obj = get<Object const>(j_column);
Columnar<T> foreign_column = ArrayInterfaceHandler::ExtractArray<T>(column_obj);
int32_t kBlocks = common::DivRoundUp(n_rows, kThreads);
CreateCSRKernel<T><<<kBlocks, kThreads>>>(foreign_column, column_id, has_missing, missing,
dh::ToSpan(*tmp_offset), s_data);
}
dh::safe_cuda(cudaMemset(s_offsets.data(), 0, sizeof(int32_t) * (n_rows + 1)));
void SimpleCSRSource::FromDeviceColumnar(std::vector<Json> const& columns,
bool has_missing, float missing) {
auto const n_cols = columns.size();
int32_t constexpr kThreads = 256;
CountValidKernel<kThreads><<<n_cols, kThreads>>>(s_d_cols, n_rows, s_offsets);
dh::caching_device_vector<int32_t> d_flag;
if (!common::CheckNAN(missing)) {
d_flag.resize(1);
thrust::fill(d_flag.begin(), d_flag.end(), 0);
}
uint32_t n_rows {0};
for (size_t i = 0; i < n_cols; ++i) {
auto const& typestr = get<String const>(columns[i]["typestr"]);
DISPATCH_TYPE(CountValid, typestr,
columns, i, has_missing, missing, &(this->page_.offset), &d_flag, &n_rows);
}
// don't pay for what you don't use.
if (!common::CheckNAN(missing)) {
int32_t flag {0};
dh::safe_cuda(cudaMemcpy(&flag, d_flag.data().get(), sizeof(int32_t), cudaMemcpyDeviceToHost));
CHECK_EQ(flag, 0) << "missing value is specifed but input data contains NaN.";
}
info.num_col_ = n_cols;
info.num_row_ = n_rows;
auto s_offsets = this->page_.offset.DeviceSpan();
thrust::device_ptr<size_t> p_offsets(s_offsets.data());
CHECK_GE(s_offsets.size(), n_rows + 1);
thrust::inclusive_scan(p_offsets, p_offsets + n_rows + 1, p_offsets);
// Created for building csr matrix, where we need to change index
// after processing each column.
dh::device_vector<size_t> tmp_offset(page_.offset.Size());
thrust::copy(p_offsets, p_offsets + n_rows + 1, tmp_offset.begin());
// Created for building csr matrix, where we need to change index after processing each
// column.
dh::device_vector<size_t> tmp_offset(this->page_.offset.Size());
dh::safe_cuda(cudaMemcpy(tmp_offset.data().get(), s_offsets.data(),
s_offsets.size_bytes(), cudaMemcpyDeviceToDevice));
// We can use null_count from columnar data format, but that will add a non-standard
// entry in the array interface, also involves accumulating from all columns. Invoking
// one copy seems easier.
this->info.num_nonzero_ = tmp_offset.back();
int device = this->page_.offset.DeviceIdx();
this->page_.data.SetDevice(device);
this->page_.data.Resize(this->info.num_nonzero_);
auto s_data = this->page_.data.DeviceSpan();
int32_t kBlocks = common::DivRoundUp(n_rows, kThreads);
for (size_t col = 0; col < n_cols; ++col) {
CreateCSRKernel<<<kBlocks, kThreads>>>(d_cols[col], col, dh::ToSpan(tmp_offset), s_data);
for (size_t i = 0; i < n_cols; ++i) {
auto const& typestr = get<String const>(columns[i]["typestr"]);
DISPATCH_TYPE(CreateCSR, typestr, columns, i, n_rows,
has_missing, missing, &tmp_offset, s_data);
}
}

13
src/data/simple_csr_source.h
View File

@@ -14,6 +14,7 @@
#include <algorithm>
#include <string>
#include <vector>
#include <limits>
#include "columnar.h"
@@ -52,8 +53,11 @@ class SimpleCSRSource : public DataSource<SparsePage> {
/*!
* \brief copy content of data from foreign **GPU** columnar buffer.
* \param interfaces_str JSON representation of cuda array interfaces.
* \param has_missing Whether did users supply their own missing value.
* \param missing The missing value set by users.
*/
void CopyFrom(std::string const& cuda_interfaces_str);
void CopyFrom(std::string const& cuda_interfaces_str, bool has_missing,
bst_float missing = std::numeric_limits<float>::quiet_NaN());
/*!
* \brief Load data from binary stream.
* \param fi the pointer to load data from.
@@ -76,9 +80,12 @@ class SimpleCSRSource : public DataSource<SparsePage> {
private:
/*!
* \brief copy content of data from foreign GPU columnar buffer.
* \param cols foreign columns data buffer.
* \param columns JSON representation of array interfaces.
* \param missing specifed missing value
*/
void FromDeviceColumnar(std::vector<Columnar> cols);
void FromDeviceColumnar(std::vector<Json> const& columns,
bool has_missing = false,
float missing = std::numeric_limits<float>::quiet_NaN());
/*! \brief internal variable, used to support iterator interface */
bool at_first_{true};
};