/
xgboost4j-gpu.cu
396 lines (340 loc) · 13.3 KB
/
xgboost4j-gpu.cu
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
#include <jni.h>
#include "../../../../src/common/device_helpers.cuh"
#include "../../../../src/common/cuda_pinned_allocator.h"
#include "../../../../src/data/array_interface.h"
#include "jvm_utils.h"
#include <xgboost/c_api.h>
namespace xgboost {
namespace jni {
template <typename T, typename Alloc>
T const *RawPtr(std::vector<T, Alloc> const &data) {
return data.data();
}
template <typename T, typename Alloc> T *RawPtr(std::vector<T, Alloc> &data) {
return data.data();
}
template <typename T> T const *RawPtr(dh::device_vector<T> const &data) {
return data.data().get();
}
template <typename T> T *RawPtr(dh::device_vector<T> &data) {
return data.data().get();
}
template <typename T> T CheckJvmCall(T const &v, JNIEnv *jenv) {
if (!v) {
CHECK(jenv->ExceptionOccurred());
jenv->ExceptionDescribe();
}
return v;
}
template <typename VCont>
void CopyColumnMask(xgboost::ArrayInterface<1> const &interface,
std::vector<Json> const &columns, cudaMemcpyKind kind,
size_t c, VCont *p_mask, Json *p_out, cudaStream_t stream) {
auto &mask = *p_mask;
auto &out = *p_out;
auto size = sizeof(typename VCont::value_type) * interface.n;
mask.resize(size);
CHECK(RawPtr(mask));
CHECK(size);
CHECK(interface.valid.Data());
dh::safe_cuda(
cudaMemcpyAsync(RawPtr(mask), interface.valid.Data(), size, kind, stream));
auto const &mask_column = columns[c]["mask"];
out["mask"] = Object();
std::vector<Json> mask_data{
Json{reinterpret_cast<Integer::Int>(RawPtr(mask))},
Json{get<Boolean const>(mask_column["data"][1])}};
out["mask"]["data"] = Array(std::move(mask_data));
if (get<Array const>(mask_column["shape"]).size() == 2) {
std::vector<Json> mask_shape{
Json{get<Integer const>(mask_column["shape"][0])},
Json{get<Integer const>(mask_column["shape"][1])}};
out["mask"]["shape"] = Array(std::move(mask_shape));
} else if (get<Array const>(mask_column["shape"]).size() == 1) {
std::vector<Json> mask_shape{
Json{get<Integer const>(mask_column["shape"][0])}};
out["mask"]["shape"] = Array(std::move(mask_shape));
} else {
LOG(FATAL) << "Invalid shape of mask";
}
out["mask"]["typestr"] = String("<t1");
out["mask"]["version"] = Integer(3);
}
template <typename DCont, typename VCont>
void CopyInterface(std::vector<xgboost::ArrayInterface<1>> &interface_arr,
std::vector<Json> const &columns, cudaMemcpyKind kind,
std::vector<DCont> *p_data, std::vector<VCont> *p_mask,
std::vector<xgboost::Json> *p_out, cudaStream_t stream) {
p_data->resize(interface_arr.size());
p_mask->resize(interface_arr.size());
p_out->resize(interface_arr.size());
for (size_t c = 0; c < interface_arr.size(); ++c) {
auto &interface = interface_arr.at(c);
size_t element_size = interface.ElementSize();
size_t size = element_size * interface.n;
auto &data = (*p_data)[c];
auto &mask = (*p_mask)[c];
data.resize(size);
dh::safe_cuda(cudaMemcpyAsync(RawPtr(data), interface.data, size, kind, stream));
auto &out = (*p_out)[c];
out = Object();
std::vector<Json> j_data{
Json{Integer(reinterpret_cast<Integer::Int>(RawPtr(data)))},
Json{Boolean{false}}};
out["data"] = Array(std::move(j_data));
out["shape"] = Array(std::vector<Json>{Json(Integer(interface.Shape(0)))});
if (interface.valid.Data()) {
CopyColumnMask(interface, columns, kind, c, &mask, &out, stream);
}
out["typestr"] = String("<f4");
out["version"] = Integer(3);
}
}
void CopyMetaInfo(Json *p_interface, dh::device_vector<float> *out, cudaStream_t stream) {
auto &j_interface = *p_interface;
CHECK_EQ(get<Array const>(j_interface).size(), 1);
auto object = get<Object>(get<Array>(j_interface)[0]);
ArrayInterface<1> interface(object);
out->resize(interface.Shape(0));
size_t element_size = interface.ElementSize();
size_t size = element_size * interface.n;
dh::safe_cuda(cudaMemcpyAsync(RawPtr(*out), interface.data, size,
cudaMemcpyDeviceToDevice, stream));
j_interface[0]["data"][0] = reinterpret_cast<Integer::Int>(RawPtr(*out));
}
template <typename DCont, typename VCont> struct DataFrame {
std::vector<DCont> data;
std::vector<VCont> valid;
std::vector<Json> interfaces;
};
class DataIteratorProxy {
DMatrixHandle proxy_;
JNIEnv *jenv_;
int jni_status_;
jobject jiter_;
bool cache_on_host_{true}; // TODO(Bobby): Make this optional.
template <typename T>
using Alloc = xgboost::common::cuda::pinned_allocator<T>;
template <typename U>
using HostVector = std::vector<U, Alloc<U>>;
// This vector is created for staging device data on host to save GPU memory.
// When space is not of concern, we can stage them on device memory directly.
std::vector<
std::unique_ptr<DataFrame<HostVector<char>, HostVector<std::uint8_t>>>>
host_columns_;
// TODO(Bobby): Use this instead of `host_columns_` if staging is not
// required.
std::vector<std::unique_ptr<DataFrame<dh::device_vector<char>,
dh::device_vector<std::uint8_t>>>>
device_columns_;
// Staging area for metainfo.
// TODO(Bobby): label_upper_bound, label_lower_bound, group.
std::vector<std::unique_ptr<dh::device_vector<float>>> labels_;
std::vector<std::unique_ptr<dh::device_vector<float>>> weights_;
std::vector<std::unique_ptr<dh::device_vector<float>>> base_margins_;
std::vector<Json> label_interfaces_;
std::vector<Json> weight_interfaces_;
std::vector<Json> margin_interfaces_;
size_t it_{0};
size_t n_batches_{0};
bool initialized_{false};
jobject last_batch_ {nullptr};
// Temp buffer on device, each `dh::device_vector` represents a column
// from cudf.
std::vector<dh::device_vector<char>> staging_data_;
std::vector<dh::device_vector<uint8_t>> staging_mask_;
cudaStream_t copy_stream_;
public:
explicit DataIteratorProxy(jobject jiter, bool cache_on_host = true)
: jiter_{jiter}, cache_on_host_{cache_on_host} {
XGProxyDMatrixCreate(&proxy_);
jni_status_ =
GlobalJvm()->GetEnv(reinterpret_cast<void **>(&jenv_), JNI_VERSION_1_6);
this->Reset();
dh::safe_cuda(cudaStreamCreateWithFlags(©_stream_, cudaStreamNonBlocking));
}
~DataIteratorProxy() { XGDMatrixFree(proxy_);
dh::safe_cuda(cudaStreamDestroy(copy_stream_));
}
DMatrixHandle GetDMatrixHandle() const { return proxy_; }
// Helper function for staging meta info.
void StageMetaInfo(Json json_interface) {
CHECK(!IsA<Null>(json_interface));
auto json_map = get<Object const>(json_interface);
if (json_map.find("label_str") == json_map.cend()) {
LOG(FATAL) << "Must have a label field.";
}
Json label = json_interface["label_str"];
CHECK(!IsA<Null>(label));
labels_.emplace_back(new dh::device_vector<float>);
CopyMetaInfo(&label, labels_.back().get(), copy_stream_);
label_interfaces_.emplace_back(label);
std::string str;
Json::Dump(label, &str);
XGDMatrixSetInfoFromInterface(proxy_, "label", str.c_str());
if (json_map.find("weight_str") != json_map.cend()) {
Json weight = json_interface["weight_str"];
CHECK(!IsA<Null>(weight));
weights_.emplace_back(new dh::device_vector<float>);
CopyMetaInfo(&weight, weights_.back().get(), copy_stream_);
weight_interfaces_.emplace_back(weight);
Json::Dump(weight, &str);
XGDMatrixSetInfoFromInterface(proxy_, "weight", str.c_str());
}
if (json_map.find("basemargin_str") != json_map.cend()) {
Json basemargin = json_interface["basemargin_str"];
base_margins_.emplace_back(new dh::device_vector<float>);
CopyMetaInfo(&basemargin, base_margins_.back().get(), copy_stream_);
margin_interfaces_.emplace_back(basemargin);
Json::Dump(basemargin, &str);
XGDMatrixSetInfoFromInterface(proxy_, "base_margin", str.c_str());
}
}
void CloseJvmBatch() {
if (last_batch_) {
jclass batch_class = CheckJvmCall(jenv_->GetObjectClass(last_batch_), jenv_);
jmethodID closeMethod = CheckJvmCall(jenv_->GetMethodID(batch_class, "close", "()V"), jenv_);
jenv_->CallVoidMethod(last_batch_, closeMethod);
last_batch_ = nullptr;
}
}
void Reset() {
it_ = 0;
this->CloseJvmBatch();
}
int32_t PullIterFromJVM() {
jclass iterClass = jenv_->FindClass("java/util/Iterator");
this->CloseJvmBatch();
jmethodID has_next =
CheckJvmCall(jenv_->GetMethodID(iterClass, "hasNext", "()Z"), jenv_);
jmethodID next = CheckJvmCall(
jenv_->GetMethodID(iterClass, "next", "()Ljava/lang/Object;"), jenv_);
if (jenv_->CallBooleanMethod(jiter_, has_next)) {
// batch should be ColumnBatch from jvm
jobject batch = CheckJvmCall(jenv_->CallObjectMethod(jiter_, next), jenv_);
jclass batch_class = CheckJvmCall(jenv_->GetObjectClass(batch), jenv_);
jmethodID getArrayInterfaceJson = CheckJvmCall(jenv_->GetMethodID(
batch_class, "getArrayInterfaceJson", "()Ljava/lang/String;"), jenv_);
auto jinterface =
static_cast<jstring>(jenv_->CallObjectMethod(batch, getArrayInterfaceJson));
CheckJvmCall(jinterface, jenv_);
char const *c_interface_str =
CheckJvmCall(jenv_->GetStringUTFChars(jinterface, nullptr), jenv_);
StageData(c_interface_str);
jenv_->ReleaseStringUTFChars(jinterface, c_interface_str);
last_batch_ = batch;
return 1;
} else {
return 0;
}
}
void StageData(std::string interface_str) {
++n_batches_;
// DataFrame
using T = decltype(host_columns_)::value_type::element_type;
host_columns_.emplace_back(std::unique_ptr<T>(new T));
// Stage the meta info.
auto json_interface =
Json::Load({interface_str.c_str(), interface_str.size()});
CHECK(!IsA<Null>(json_interface));
StageMetaInfo(json_interface);
Json features = json_interface["features_str"];
auto json_columns = get<Array const>(features);
std::vector<ArrayInterface<1>> interfaces;
// Stage the data
for (auto &json_col : json_columns) {
auto column = ArrayInterface<1>(get<Object const>(json_col));
interfaces.emplace_back(column);
}
Json::Dump(features, &interface_str);
CopyInterface(interfaces, json_columns, cudaMemcpyDeviceToHost,
&host_columns_.back()->data, &host_columns_.back()->valid,
&host_columns_.back()->interfaces, copy_stream_);
XGProxyDMatrixSetDataCudaColumnar(proxy_, interface_str.c_str());
it_++;
}
int NextFirstLoop() {
try {
dh::safe_cuda(cudaStreamSynchronize(copy_stream_));
if (this->PullIterFromJVM()) {
return 1;
} else {
initialized_ = true;
return 0;
}
} catch (dmlc::Error const &e) {
if (jni_status_ == JNI_EDETACHED) {
GlobalJvm()->DetachCurrentThread();
}
LOG(FATAL) << e.what();
}
LOG(FATAL) << "Unreachable";
return 1;
}
int NextSecondLoop() {
std::string str;
// Meta
auto const &label = this->label_interfaces_.at(it_);
Json::Dump(label, &str);
XGDMatrixSetInfoFromInterface(proxy_, "label", str.c_str());
if (n_batches_ == this->weight_interfaces_.size()) {
auto const &weight = this->weight_interfaces_.at(it_);
Json::Dump(weight, &str);
XGDMatrixSetInfoFromInterface(proxy_, "weight", str.c_str());
}
if (n_batches_ == this->margin_interfaces_.size()) {
auto const &base_margin = this->margin_interfaces_.at(it_);
Json::Dump(base_margin, &str);
XGDMatrixSetInfoFromInterface(proxy_, "base_margin", str.c_str());
}
// Data
auto const &json_interface = host_columns_.at(it_)->interfaces;
std::vector<ArrayInterface<1>> in;
for (auto interface : json_interface) {
auto column = ArrayInterface<1>(get<Object const>(interface));
in.emplace_back(column);
}
std::vector<Json> out;
CopyInterface(in, json_interface, cudaMemcpyHostToDevice, &staging_data_,
&staging_mask_, &out, nullptr);
Json temp{Array(std::move(out))};
std::string interface_str;
Json::Dump(temp, &interface_str);
XGProxyDMatrixSetDataCudaColumnar(proxy_, interface_str.c_str());
it_++;
return 1;
}
int Next() {
if (!initialized_) {
return NextFirstLoop();
} else {
if (it_ == n_batches_) {
return 0;
}
return NextSecondLoop();
}
};
};
namespace {
void Reset(DataIterHandle self) {
static_cast<xgboost::jni::DataIteratorProxy *>(self)->Reset();
}
int Next(DataIterHandle self) {
return static_cast<xgboost::jni::DataIteratorProxy *>(self)->Next();
}
} // anonymous namespace
XGB_DLL jint XGDeviceQuantileDMatrixCreateFromCallbackImpl(JNIEnv *jenv, jclass jcls,
jobject jiter,
jfloat jmissing,
jint jmax_bin, jint jnthread,
jlongArray jout) {
xgboost::jni::DataIteratorProxy proxy(jiter);
DMatrixHandle result;
auto ret = XGDeviceQuantileDMatrixCreateFromCallback(
&proxy, proxy.GetDMatrixHandle(), Reset, Next, jmissing, jnthread,
jmax_bin, &result);
setHandle(jenv, jout, result);
return ret;
}
} // namespace jni
} // namespace xgboost