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

Remove internal use of gpu_id. (#9568)

Browse Source
This commit is contained in:
Jiaming Yuan
2023-09-20 23:29:51 +08:00
committed by GitHub
parent 38ac52dd87
commit 8c676c889d
121 changed files with 1012 additions and 1044 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
src/metric/auc.cc
View File

@@ -23,8 +23,7 @@
#include "xgboost/linalg.h"
#include "xgboost/metric.h"
namespace xgboost {
namespace metric {
namespace xgboost::metric {
// tag the this file, used by force static link later.
DMLC_REGISTRY_FILE_TAG(auc);
/**
@@ -257,10 +256,10 @@ template <typename Curve>
class EvalAUC : public MetricNoCache {
double Eval(const HostDeviceVector<bst_float> &preds, const MetaInfo &info) override {
double auc {0};
if (ctx_->gpu_id != Context::kCpuId) {
preds.SetDevice(ctx_->gpu_id);
info.labels.SetDevice(ctx_->gpu_id);
info.weights_.SetDevice(ctx_->gpu_id);
if (ctx_->Device().IsCUDA()) {
preds.SetDevice(ctx_->Device());
info.labels.SetDevice(ctx_->Device());
info.weights_.SetDevice(ctx_->Device());
}
// We use the global size to handle empty dataset.
std::array<size_t, 2> meta{info.labels.Size(), preds.Size()};
@@ -329,7 +328,7 @@ class EvalROCAUC : public EvalAUC<EvalROCAUC> {
double auc{0};
uint32_t valid_groups = 0;
auto n_threads = ctx_->Threads();
if (ctx_->gpu_id == Context::kCpuId) {
if (ctx_->IsCPU()) {
std::tie(auc, valid_groups) =
RankingAUC<true>(ctx_, predts.ConstHostVector(), info, n_threads);
} else {
@@ -344,7 +343,7 @@ class EvalROCAUC : public EvalAUC<EvalROCAUC> {
double auc{0};
auto n_threads = ctx_->Threads();
CHECK_NE(n_classes, 0);
if (ctx_->gpu_id == Context::kCpuId) {
if (ctx_->IsCPU()) {
auc = MultiClassOVR(ctx_, predts.ConstHostVector(), info, n_classes, n_threads, BinaryROCAUC);
} else {
auc = GPUMultiClassROCAUC(ctx_, predts.ConstDeviceSpan(), info, &this->d_cache_, n_classes);
@@ -355,7 +354,7 @@ class EvalROCAUC : public EvalAUC<EvalROCAUC> {
std::tuple<double, double, double>
EvalBinary(HostDeviceVector<float> const &predts, MetaInfo const &info) {
double fp, tp, auc;
if (ctx_->gpu_id == Context::kCpuId) {
if (ctx_->IsCPU()) {
std::tie(fp, tp, auc) = BinaryROCAUC(ctx_, predts.ConstHostVector(),
info.labels.HostView().Slice(linalg::All(), 0),
common::OptionalWeights{info.weights_.ConstHostSpan()});
@@ -367,7 +366,7 @@ class EvalROCAUC : public EvalAUC<EvalROCAUC> {
}
public:
char const* Name() const override {
[[nodiscard]] char const* Name() const override {
return "auc";
}
};
@@ -405,7 +404,7 @@ class EvalPRAUC : public EvalAUC<EvalPRAUC> {
std::tuple<double, double, double>
EvalBinary(HostDeviceVector<float> const &predts, MetaInfo const &info) {
double pr, re, auc;
if (ctx_->gpu_id == Context::kCpuId) {
if (ctx_->IsCPU()) {
std::tie(pr, re, auc) =
BinaryPRAUC(ctx_, predts.ConstHostSpan(), info.labels.HostView().Slice(linalg::All(), 0),
common::OptionalWeights{info.weights_.ConstHostSpan()});
@@ -418,7 +417,7 @@ class EvalPRAUC : public EvalAUC<EvalPRAUC> {
double EvalMultiClass(HostDeviceVector<float> const &predts, MetaInfo const &info,
size_t n_classes) {
if (ctx_->gpu_id == Context::kCpuId) {
if (ctx_->IsCPU()) {
auto n_threads = this->ctx_->Threads();
return MultiClassOVR(ctx_, predts.ConstHostSpan(), info, n_classes, n_threads, BinaryPRAUC);
} else {
@@ -431,7 +430,7 @@ class EvalPRAUC : public EvalAUC<EvalPRAUC> {
double auc{0};
uint32_t valid_groups = 0;
auto n_threads = ctx_->Threads();
if (ctx_->gpu_id == Context::kCpuId) {
if (ctx_->IsCPU()) {
auto labels = info.labels.Data()->ConstHostSpan();
if (std::any_of(labels.cbegin(), labels.cend(), PRAUCLabelInvalid{})) {
InvalidLabels();
@@ -446,7 +445,7 @@ class EvalPRAUC : public EvalAUC<EvalPRAUC> {
}
public:
const char *Name() const override { return "aucpr"; }
[[nodiscard]] const char *Name() const override { return "aucpr"; }
};
XGBOOST_REGISTER_METRIC(AUCPR, "aucpr")
@@ -473,5 +472,4 @@ std::pair<double, std::uint32_t> GPURankingPRAUC(Context const *, common::Span<f
return {};
}
#endif
} // namespace metric
} // namespace xgboost
} // namespace xgboost::metric

2
src/metric/auc.cu
View File

@@ -824,7 +824,7 @@ std::pair<double, std::uint32_t> GPURankingPRAUC(Context const *ctx,
common::Span<float const> predts,
MetaInfo const &info,
std::shared_ptr<DeviceAUCCache> *p_cache) {
dh::safe_cuda(cudaSetDevice(ctx->gpu_id));
dh::safe_cuda(cudaSetDevice(ctx->Ordinal()));
if (predts.empty()) {
return std::make_pair(0.0, static_cast<uint32_t>(0));
}

14
src/metric/multiclass_metric.cu
View File

@@ -127,24 +127,24 @@ class MultiClassMetricsReduction {
#endif // XGBOOST_USE_CUDA
PackedReduceResult Reduce(const Context& tparam, int device, size_t n_class,
PackedReduceResult Reduce(const Context& ctx, DeviceOrd device, size_t n_class,
const HostDeviceVector<bst_float>& weights,
const HostDeviceVector<bst_float>& labels,
const HostDeviceVector<bst_float>& preds) {
PackedReduceResult result;
if (device < 0) {
if (device.IsCPU()) {
result =
CpuReduceMetrics(weights, labels, preds, n_class, tparam.Threads());
CpuReduceMetrics(weights, labels, preds, n_class, ctx.Threads());
}
#if defined(XGBOOST_USE_CUDA)
else { // NOLINT
device_ = tparam.gpu_id;
device_ = ctx.Device();
preds.SetDevice(device_);
labels.SetDevice(device_);
weights.SetDevice(device_);
dh::safe_cuda(cudaSetDevice(device_));
dh::safe_cuda(cudaSetDevice(device_.ordinal));
result = DeviceReduceMetrics(weights, labels, preds, n_class);
}
#endif // defined(XGBOOST_USE_CUDA)
@@ -154,7 +154,7 @@ class MultiClassMetricsReduction {
private:
#if defined(XGBOOST_USE_CUDA)
dh::PinnedMemory label_error_;
int device_{-1};
DeviceOrd device_{DeviceOrd::CPU()};
#endif // defined(XGBOOST_USE_CUDA)
};
@@ -176,7 +176,7 @@ struct EvalMClassBase : public MetricNoCache {
CHECK_GE(nclass, 1U)
<< "mlogloss and merror are only used for multi-class classification,"
<< " use logloss for binary classification";
int device = ctx_->gpu_id;
auto device = ctx_->Device();
auto result =
reducer_.Reduce(*ctx_, device, nclass, info.weights_, *info.labels.Data(), preds);
dat[0] = result.Residue();

4
src/metric/rank_metric.cu
View File

@@ -35,7 +35,7 @@ PackedReduceResult PreScore(Context const *ctx, MetaInfo const &info,
auto d_gptr = p_cache->DataGroupPtr(ctx);
auto d_label = info.labels.View(ctx->Device()).Slice(linalg::All(), 0);
predt.SetDevice(ctx->gpu_id);
predt.SetDevice(ctx->Device());
auto d_rank_idx = p_cache->SortedIdx(ctx, predt.ConstDeviceSpan());
auto topk = p_cache->Param().TopK();
auto d_weight = common::MakeOptionalWeights(ctx, info.weights_);
@@ -90,7 +90,7 @@ PackedReduceResult NDCGScore(Context const *ctx, MetaInfo const &info,
CHECK_EQ(d_weight.weights.size(), p_cache->Groups());
}
auto d_label = info.labels.View(ctx->Device()).Slice(linalg::All(), 0);
predt.SetDevice(ctx->gpu_id);
predt.SetDevice(ctx->Device());
auto d_predt = linalg::MakeTensorView(ctx, predt.ConstDeviceSpan(), predt.Size());
auto d_group_ptr = p_cache->DataGroupPtr(ctx);

12
src/metric/survival_metric.cu
View File

@@ -130,18 +130,18 @@ class ElementWiseSurvivalMetricsReduction {
const HostDeviceVector<bst_float>& preds) {
PackedReduceResult result;
if (ctx.gpu_id < 0) {
if (ctx.IsCPU()) {
result = CpuReduceMetrics(weights, labels_lower_bound, labels_upper_bound,
preds, ctx.Threads());
}
#if defined(XGBOOST_USE_CUDA)
else { // NOLINT
preds.SetDevice(ctx.gpu_id);
labels_lower_bound.SetDevice(ctx.gpu_id);
labels_upper_bound.SetDevice(ctx.gpu_id);
weights.SetDevice(ctx.gpu_id);
preds.SetDevice(ctx.Device());
labels_lower_bound.SetDevice(ctx.Device());
labels_upper_bound.SetDevice(ctx.Device());
weights.SetDevice(ctx.Device());
dh::safe_cuda(cudaSetDevice(ctx.gpu_id));
dh::safe_cuda(cudaSetDevice(ctx.Ordinal()));
result = DeviceReduceMetrics(weights, labels_lower_bound, labels_upper_bound, preds);
}
#endif // defined(XGBOOST_USE_CUDA)