Merge branch 'develop' of github.com:baidu/Paddle into feature/backward

cblas_new
Yu Yang 8 years ago
commit bc146e8f6d

@ -7,17 +7,8 @@ INCLUDE_DIRECTORIES(${EIGEN_SOURCE_DIR}/src/extern_eigen3)
ExternalProject_Add(
extern_eigen3
${EXTERNAL_PROJECT_LOG_ARGS}
# for latest version, please get from official website
# URL "https://bitbucket.org/eigen/eigen/get/3.3.4.tar.gz"
# URL_MD5 "1a47e78efe365a97de0c022d127607c3"
# for no-ssl http support, please get from bazel's mirror
# URL "http://mirror.bazel.build/bitbucket.org/eigen/eigen/get/f3a22f35b044.tar.gz"
# URL_MD5 "4645c66075982da6fa0bcf6b20f3e8f7"
# get from github mirror
GIT_REPOSITORY "https://github.com/RLovelett/eigen.git"
GIT_TAG "a46d2e7337c4656f00abe54a8115f6d76153a048"
GIT_TAG "master"
PREFIX ${EIGEN_SOURCE_DIR}
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""

@ -153,7 +153,7 @@ set(CUDA_PROPAGATE_HOST_FLAGS OFF)
# Release/Debug flags set by cmake. Such as -O3 -g -DNDEBUG etc.
# So, don't set these flags here.
LIST(APPEND CUDA_NVCC_FLAGS -std=c++11)
LIST(APPEND CUDA_NVCC_FLAGS -std=c++11 --default-stream per-thread)
LIST(APPEND CUDA_NVCC_FLAGS --use_fast_math)
if(CMAKE_BUILD_TYPE STREQUAL "Debug")

@ -19,6 +19,8 @@ import (
"net"
"net/http"
"net/rpc"
"os"
"os/signal"
"strconv"
"strings"
"time"
@ -68,6 +70,20 @@ func main() {
store = &master.InMemStore{}
}
shutdown := func() {
log.Infoln("shutting down gracefully")
err := store.Shutdown()
if err != nil {
log.Errorln(err)
}
}
// Guaranteed to run even panic happens.
defer shutdown()
c := make(chan os.Signal, 1)
signal.Notify(c, os.Interrupt)
s, err := master.NewService(store, *chunkPerTask, *taskTimeoutDur, *taskTimeoutMax)
if err != nil {
log.Fatal(err)
@ -84,8 +100,12 @@ func main() {
log.Fatal(err)
}
err = http.Serve(l, nil)
if err != nil {
log.Fatal(err)
}
go func() {
err = http.Serve(l, nil)
if err != nil {
log.Fatal(err)
}
}()
<-c
}

@ -18,6 +18,8 @@ import (
"net"
"net/http"
"net/rpc"
"os"
"os/signal"
"strconv"
"time"
@ -33,7 +35,8 @@ func main() {
index := flag.Int("index", -1, "index of this pserver, should be larger or equal than 0")
etcdEndpoint := flag.String("etcd-endpoint", "http://127.0.0.1:2379",
"comma separated endpoint string for pserver to connect to etcd")
etcdTimeout := flag.Duration("etcd-timeout", 5*time.Second, "timeout for etcd calls")
dialTimeout := flag.Duration("dial-timeout", 5*time.Second, "dial timeout")
etcdTTL := flag.Int("etcd-ttl", 5, "etcd time to live in seconds")
numPservers := flag.Int("num-pservers", 1, "total pserver count in a training job")
checkpointPath := flag.String("checkpoint-path", "/checkpoints/", "save checkpoint path")
checkpointInterval := flag.Duration("checkpoint-interval", 600*time.Second, "save checkpoint per interval seconds")
@ -53,7 +56,7 @@ func main() {
if *index >= 0 {
idx = *index
} else {
e = pserver.NewEtcdClient(*etcdEndpoint, *numPservers, *etcdTimeout)
e = pserver.NewEtcdClient(*etcdEndpoint, *numPservers, *dialTimeout, *etcdTTL)
idx, err = e.Register(*port)
candy.Must(err)
@ -67,6 +70,20 @@ func main() {
}
}
shutdown := func() {
log.Infoln("shutting down gracefully")
sErr := e.Shutdown()
if sErr != nil {
log.Errorln(sErr)
}
}
// Guaranteed to run even panic happens.
defer shutdown()
c := make(chan os.Signal, 1)
signal.Notify(c, os.Interrupt)
s, err := pserver.NewService(idx, *checkpointInterval, *checkpointPath, e, cp)
candy.Must(err)
@ -77,7 +94,11 @@ func main() {
l, err := net.Listen("tcp", ":"+strconv.Itoa(*port))
candy.Must(err)
log.Infof("start pserver at port %d", *port)
err = http.Serve(l, nil)
candy.Must(err)
go func() {
log.Infof("start pserver at port %d", *port)
err = http.Serve(l, nil)
candy.Must(err)
}()
<-c
}

176
go/glide.lock generated

@ -1,15 +1,105 @@
hash: a8faea3a363468a88917ddeb3b1c9ea36886fb2c622acbad42604fa9cb4d3855
updated: 2017-07-11T10:04:40.786745417+08:00
hash: 2a1c0eca5c07a130e3d224f9821f96cfa37a39bf6bce141c855bbc57ef569f1c
updated: 2017-07-29T07:34:48.722757905+08:00
imports:
- name: github.com/beorn7/perks
version: 4c0e84591b9aa9e6dcfdf3e020114cd81f89d5f9
subpackages:
- quantile
- name: github.com/boltdb/bolt
version: 583e8937c61f1af6513608ccc75c97b6abdf4ff9
- name: github.com/cockroachdb/cmux
version: 112f0506e7743d64a6eb8fedbcff13d9979bbf92
- name: github.com/coreos/etcd
version: cb2a496c4ddd1c87a9f280e116649b599999ec79
version: c31bec0f29facff13f7c3e3d948e55dd6689ed42
subpackages:
- alarm
- auth
- auth/authpb
- client
- clientv3
- clientv3/concurrency
- compactor
- discovery
- embed
- error
- etcdserver
- etcdserver/api
- etcdserver/api/v2http
- etcdserver/api/v2http/httptypes
- etcdserver/api/v3client
- etcdserver/api/v3election
- etcdserver/api/v3election/v3electionpb
- etcdserver/api/v3election/v3electionpb/gw
- etcdserver/api/v3lock
- etcdserver/api/v3lock/v3lockpb
- etcdserver/api/v3lock/v3lockpb/gw
- etcdserver/api/v3rpc
- etcdserver/api/v3rpc/rpctypes
- etcdserver/auth
- etcdserver/etcdserverpb
- etcdserver/etcdserverpb/gw
- etcdserver/membership
- etcdserver/stats
- lease
- lease/leasehttp
- lease/leasepb
- mvcc
- mvcc/backend
- mvcc/mvccpb
- pkg/adt
- pkg/contention
- pkg/cors
- pkg/cpuutil
- pkg/crc
- pkg/debugutil
- pkg/fileutil
- pkg/httputil
- pkg/idutil
- pkg/ioutil
- pkg/logutil
- pkg/monotime
- pkg/netutil
- pkg/pathutil
- pkg/pbutil
- pkg/runtime
- pkg/schedule
- pkg/srv
- pkg/tlsutil
- pkg/transport
- pkg/types
- pkg/wait
- proxy/grpcproxy/adapter
- raft
- raft/raftpb
- rafthttp
- snap
- snap/snappb
- store
- version
- wal
- wal/walpb
- name: github.com/coreos/go-semver
version: 8ab6407b697782a06568d4b7f1db25550ec2e4c6
subpackages:
- semver
- name: github.com/coreos/go-systemd
version: 48702e0da86bd25e76cfef347e2adeb434a0d0a6
subpackages:
- daemon
- journal
- util
- name: github.com/coreos/pkg
version: 3ac0863d7acf3bc44daf49afef8919af12f704ef
subpackages:
- capnslog
- name: github.com/dgrijalva/jwt-go
version: d2709f9f1f31ebcda9651b03077758c1f3a0018c
- name: github.com/ghodss/yaml
version: 0ca9ea5df5451ffdf184b4428c902747c2c11cd7
- name: github.com/gogo/protobuf
version: 909568be09de550ed094403c2bf8a261b5bb730a
subpackages:
- proto
- name: github.com/golang/protobuf
version: 4bd1920723d7b7c925de087aa32e2187708897f7
subpackages:
@ -17,14 +107,61 @@ imports:
- proto
- name: github.com/golang/snappy
version: 553a641470496b2327abcac10b36396bd98e45c9
- name: github.com/google/btree
version: 925471ac9e2131377a91e1595defec898166fe49
- name: github.com/grpc-ecosystem/go-grpc-prometheus
version: 6b7015e65d366bf3f19b2b2a000a831940f0f7e0
- name: github.com/grpc-ecosystem/grpc-gateway
version: 18d159699f2e83fc5bb9ef2f79465ca3f3122676
subpackages:
- runtime
- runtime/internal
- utilities
- name: github.com/jonboulle/clockwork
version: 2eee05ed794112d45db504eb05aa693efd2b8b09
- name: github.com/matttproud/golang_protobuf_extensions
version: c12348ce28de40eed0136aa2b644d0ee0650e56c
subpackages:
- pbutil
- name: github.com/namsral/flag
version: 71ceffbeb0ba60fccc853971bb3ed4d7d90bfd04
- name: github.com/PaddlePaddle/recordio
version: edfb82af0739c84f241c87390ec5649c7b28c129
version: 0432dee9fd4b24fb6840fb20a8c055b0c933fb81
- name: github.com/prometheus/client_golang
version: c5b7fccd204277076155f10851dad72b76a49317
subpackages:
- prometheus
- name: github.com/prometheus/client_model
version: 6f3806018612930941127f2a7c6c453ba2c527d2
subpackages:
- go
- name: github.com/prometheus/common
version: 49fee292b27bfff7f354ee0f64e1bc4850462edf
subpackages:
- expfmt
- internal/bitbucket.org/ww/goautoneg
- model
- name: github.com/prometheus/procfs
version: a1dba9ce8baed984a2495b658c82687f8157b98f
subpackages:
- xfs
- name: github.com/sirupsen/logrus
version: 7f976d3a76720c4c27af2ba716b85d2e0a7e38b1
version: a3f95b5c423586578a4e099b11a46c2479628cac
- name: github.com/topicai/candy
version: 1b9030d056fa9f8c4b1f9c91b52fe4b8ab4cd8cc
- name: github.com/ugorji/go
version: ded73eae5db7e7a0ef6f55aace87a2873c5d2b74
subpackages:
- codec
- name: github.com/xiang90/probing
version: 07dd2e8dfe18522e9c447ba95f2fe95262f63bb2
- name: golang.org/x/crypto
version: 1351f936d976c60a0a48d728281922cf63eafb8d
repo: https://github.com/golang/crypto.git
vcs: git
subpackages:
- bcrypt
- blowfish
- name: golang.org/x/net
version: c8c74377599bd978aee1cf3b9b63a8634051cec2
subpackages:
@ -36,11 +173,15 @@ imports:
- lex/httplex
- trace
- name: golang.org/x/sys
version: abf9c25f54453410d0c6668e519582a9e1115027
version: 0f826bdd13b500be0f1d4004938ad978fcc6031e
repo: https://github.com/golang/sys.git
vcs: git
subpackages:
- unix
- name: golang.org/x/text
version: cfdf022e86b4ecfb646e1efbd7db175dd623a8fa
version: 836efe42bb4aa16aaa17b9c155d8813d336ed720
repo: https://github.com/golang/text.git
vcs: git
subpackages:
- secure/bidirule
- transform
@ -60,4 +201,23 @@ imports:
- stats
- tap
- transport
testImports: []
- name: gopkg.in/yaml.v2
version: cd8b52f8269e0feb286dfeef29f8fe4d5b397e0b
testImports:
- name: github.com/davecgh/go-spew
version: 04cdfd42973bb9c8589fd6a731800cf222fde1a9
subpackages:
- spew
- name: github.com/docker/docker
version: b6d164e6c46d8115b146e4c3ac93784e9ef8b49e
subpackages:
- pkg/ioutils
- pkg/longpath
- name: github.com/pmezard/go-difflib
version: d8ed2627bdf02c080bf22230dbb337003b7aba2d
subpackages:
- difflib
- name: github.com/stretchr/testify
version: 05e8a0eda380579888eb53c394909df027f06991
subpackages:
- assert

@ -6,8 +6,19 @@ import:
subpackages:
- clientv3
- clientv3/concurrency
- embed
- etcdserver
- package: github.com/namsral/flag
version: ^1.7.4-pre
- package: github.com/sirupsen/logrus
version: ^1.0.0
- package: github.com/topicai/candy
- package: golang.org/x/crypto
vcs: git
repo: https://github.com/golang/crypto.git
- package: golang.org/x/sys
vcs: git
repo: https://github.com/golang/sys.git
- package: golang.org/x/text
vcs: git
repo: https://github.com/golang/text.git

@ -39,15 +39,12 @@ type EtcdClient struct {
statePath string
client *clientv3.Client
lock *concurrency.Mutex
sess *concurrency.Session
}
// NewEtcdClient creates a new EtcdClient.
func NewEtcdClient(endpoints []string, addr string, lockPath, addrPath, statePath string, ttlSec int) (*EtcdClient, error) {
log.Debugf("Connecting to etcd at %v", endpoints)
// TODO(helin): gracefully shutdown etcd store. Because etcd
// store holds a etcd lock, even though the lock will expire
// when the lease timeout, we need to implement graceful
// shutdown to release the lock.
cli, err := clientv3.New(clientv3.Config{
Endpoints: endpoints,
DialTimeout: dialTimeout,
@ -67,12 +64,12 @@ func NewEtcdClient(endpoints []string, addr string, lockPath, addrPath, statePat
// one master running, but split-brain problem may cause
// multiple master servers running), and the cluster management
// software will kill one of them.
log.Debugf("Trying to acquire lock at %s.", lockPath)
log.Infof("Trying to acquire lock at %s.", lockPath)
err = lock.Lock(context.TODO())
if err != nil {
return nil, err
}
log.Debugf("Successfully acquired lock at %s.", lockPath)
log.Infof("Successfully acquired lock at %s.", lockPath)
put := clientv3.OpPut(addrPath, addr)
resp, err := cli.Txn(context.Background()).If(lock.IsOwner()).Then(put).Commit()
@ -89,6 +86,7 @@ func NewEtcdClient(endpoints []string, addr string, lockPath, addrPath, statePat
statePath: statePath,
client: cli,
lock: lock,
sess: sess,
}
return e, nil
@ -157,6 +155,21 @@ func (e *EtcdClient) Load() ([]byte, error) {
return state, nil
}
// Shutdown shuts down the etcd client gracefully.
func (e *EtcdClient) Shutdown() error {
err := e.sess.Close()
newErr := e.client.Close()
if newErr != nil {
if err == nil {
err = newErr
} else {
log.Errorln(newErr)
}
}
return err
}
// GetKey gets the value by the specify key.
func GetKey(c *clientv3.Client, key string, timeout time.Duration) (string, error) {
ctx, cancel := context.WithTimeout(context.Background(), timeout)

@ -40,3 +40,8 @@ func (m *InMemStore) Load() ([]byte, error) {
return m.buf, nil
}
// Shutdown shuts down the in mem store.
func (m *InMemStore) Shutdown() error {
return nil
}

@ -50,6 +50,7 @@ var ErrPassAfter = errors.New("pass number larger than master")
type Store interface {
Save([]byte) error
Load() ([]byte, error)
Shutdown() error
}
// Chunk is a chunk of data consisted of several data instances.

@ -0,0 +1,68 @@
package master_test
import (
"os"
"testing"
"time"
"github.com/PaddlePaddle/Paddle/go/master"
"github.com/coreos/etcd/clientv3"
"github.com/coreos/etcd/embed"
"github.com/docker/docker/pkg/ioutils"
"github.com/stretchr/testify/assert"
)
func TestNewServiceWithEtcd(t *testing.T) {
// setup an embed etcd server
etcdDir, err := ioutils.TempDir("", "")
if err != nil {
t.Fatal(err)
}
cfg := embed.NewConfig()
cfg.Dir = etcdDir
e, err := embed.StartEtcd(cfg)
if err != nil {
t.Fatal(err)
}
defer func() {
e.Close()
if err := os.RemoveAll(etcdDir); err != nil {
t.Fatal(err)
}
}()
select {
case <-e.Server.ReadyNotify():
t.Log("Server is ready!")
case <-time.After(60 * time.Second):
e.Server.Stop() // trigger a shutdown
t.Fatal("Server took too long to start!")
}
ep := []string{"127.0.0.1:2379"}
masterAddr := "127.0.0.1:3306"
store, err := master.NewEtcdClient(ep, masterAddr, master.DefaultLockPath, master.DefaultAddrPath, master.DefaultStatePath, 30)
if err != nil {
t.Fatal(err)
}
_, err = master.NewService(store, 10, 10, 3)
if err != nil {
t.Fatal(err)
}
cli, err := clientv3.New(clientv3.Config{
Endpoints: ep,
DialTimeout: 3 * time.Second,
})
if err != nil {
t.Fatal(err)
}
v, err := master.GetKey(cli, master.DefaultAddrPath, 3*time.Second)
if err != nil {
t.Fatal(err)
}
if err := cli.Close(); err != nil {
t.Fatal(err)
}
// test master process registry itself into etcd server.
assert.Equal(t, masterAddr, v, "master process should registry itself into etcd server.")
}

@ -55,10 +55,10 @@ var curHandle C.paddle_pserver_client
func add(c *client.Client) C.paddle_pserver_client {
mu.Lock()
defer mu.Unlock()
client := curHandle
cli := curHandle
curHandle++
handleMap[client] = c
return client
handleMap[cli] = c
return cli
}
func get(client C.paddle_pserver_client) *client.Client {

@ -34,16 +34,19 @@ const (
PsPath = "/ps/"
// PsCheckpoint is the etcd path for store checkpoints information
PsCheckpoint = "/checkpoints/"
retryTimeout = 5 * time.Second
)
// EtcdClient is the etcd client that the pserver uses for fault
// tolerance, service registry and coordination.
type EtcdClient struct {
numPservers int
etcdEndpoints string
etcdClient *clientv3.Client
// etcdTimeout is also used as retry intervals.
etcdTimeout time.Duration
numPservers int
endpoints string
client *clientv3.Client
sess *concurrency.Session
dialTimeout time.Duration
ttlSec int
// FIXME: ensure GetExternalIP gets the correct ip for trainers to connect.
externalIP string
// desired number of pservers in the job.
@ -52,11 +55,12 @@ type EtcdClient struct {
}
// NewEtcdClient creates an EtcdClient
func NewEtcdClient(endpoints string, numPservers int, timeout time.Duration) *EtcdClient {
func NewEtcdClient(endpoints string, numPservers int, dialtimeout time.Duration, ttlSec int) *EtcdClient {
return &EtcdClient{
etcdTimeout: timeout,
numPservers: numPservers,
etcdEndpoints: endpoints,
dialTimeout: dialtimeout,
ttlSec: ttlSec,
numPservers: numPservers,
endpoints: endpoints,
}
}
@ -64,7 +68,6 @@ func NewEtcdClient(endpoints string, numPservers int, timeout time.Duration) *Et
//
// Register returns the index of the current pserver.
func (e *EtcdClient) Register(port int) (int, error) {
var err error
e.externalIP, err = networkhelper.GetExternalIP()
if err != nil {
@ -72,19 +75,26 @@ func (e *EtcdClient) Register(port int) (int, error) {
}
// initialize connection to etcd.
ep := strings.Split(e.etcdEndpoints, ",")
ep := strings.Split(e.endpoints, ",")
for {
cli, err := clientv3.New(clientv3.Config{
Endpoints: ep,
DialTimeout: e.etcdTimeout,
DialTimeout: e.dialTimeout,
})
if err != nil {
log.Errorf("connect to etcd error: %v", err)
time.Sleep(e.etcdTimeout)
time.Sleep(retryTimeout)
continue
}
e.client = cli
sess, err := concurrency.NewSession(cli, concurrency.WithTTL(e.ttlSec))
if err != nil {
log.Errorf("create etcd session error: %v", err)
time.Sleep(retryTimeout)
continue
}
e.etcdClient = cli
log.Debugf("inited client to %s", e.etcdEndpoints)
e.sess = sess
log.Debugf("inited client to %s", e.endpoints)
break
}
// init /ps_desired using transaction, for multiple pservers may want to write
@ -95,7 +105,7 @@ func (e *EtcdClient) Register(port int) (int, error) {
cancel()
if err != nil {
log.Warn(err)
time.Sleep(e.etcdTimeout)
time.Sleep(retryTimeout)
continue
}
break
@ -106,18 +116,18 @@ func (e *EtcdClient) Register(port int) (int, error) {
// wait and set s.desired init value
for {
ctx, cancel := context.WithTimeout(context.Background(), time.Second)
resp, err := e.etcdClient.Get(ctx, PsDesired)
resp, err := e.client.Get(ctx, PsDesired)
cancel()
if err != nil {
log.Errorf("getting %s error: %v", PsDesired, err)
time.Sleep(e.etcdTimeout)
time.Sleep(retryTimeout)
continue
}
if len(resp.Kvs) != 0 {
e.desired, err = strconv.Atoi(string(resp.Kvs[0].Value))
if err != nil {
log.Errorf("value of %s invalid %v\n", PsDesired, err)
time.Sleep(e.etcdTimeout)
time.Sleep(retryTimeout)
// NOTE: wait util ps_desired value change
continue
}
@ -134,7 +144,7 @@ func (e *EtcdClient) Register(port int) (int, error) {
cancel()
if err != nil {
log.Warn(err)
time.Sleep(e.etcdTimeout)
time.Sleep(retryTimeout)
continue
}
break
@ -144,10 +154,10 @@ func (e *EtcdClient) Register(port int) (int, error) {
}
func (e *EtcdClient) initDesiredPservers(ctx context.Context, numPservers int) (*clientv3.TxnResponse, error) {
return concurrency.NewSTM(e.etcdClient, func(c concurrency.STM) error {
return concurrency.NewSTM(e.client, func(c concurrency.STM) error {
dsStr := c.Get(PsDesired)
if dsStr == "" {
c.Put(PsDesired, strconv.Itoa(numPservers))
c.Put(PsDesired, strconv.Itoa(numPservers), clientv3.WithLease(e.sess.Lease()))
}
return nil
}, concurrency.WithAbortContext(ctx), concurrency.WithIsolation(concurrency.RepeatableReads))
@ -156,7 +166,7 @@ func (e *EtcdClient) initDesiredPservers(ctx context.Context, numPservers int) (
// registerPserverEtcd registers pserver node on etcd using transaction.
func (e *EtcdClient) registerPserverEtcd(ctx context.Context, port int) (int, error) {
var idx int
_, err := concurrency.NewSTM(e.etcdClient, func(c concurrency.STM) error {
_, err := concurrency.NewSTM(e.client, func(c concurrency.STM) error {
registered := false
for i := 0; i < e.desired; i++ {
psKey := PsPath + strconv.Itoa(i)
@ -165,26 +175,10 @@ func (e *EtcdClient) registerPserverEtcd(ctx context.Context, port int) (int, er
log.Debugf("got value (%s) for key: %s", ps, psKey)
if ps == "" {
resp, err := e.etcdClient.Grant(context.TODO(), 5)
if err != nil {
log.Fatal(err)
}
// find the first id and write info
pserverAddr := e.externalIP + ":" + strconv.Itoa(port)
c.Put(psKey, pserverAddr, clientv3.WithLease(resp.ID))
c.Put(psKey, pserverAddr, clientv3.WithLease(e.sess.Lease()))
log.Debugf("set pserver node %s with value %s", psKey, pserverAddr)
ch, kaerr := e.etcdClient.KeepAlive(context.TODO(), resp.ID)
if kaerr != nil {
log.Errorf("keepalive etcd node error: %v", kaerr)
return kaerr
}
// Eat the keep alive message so etcd
// will not expire the lease.
go func(ch <-chan *clientv3.LeaseKeepAliveResponse) {
ka := <-ch
log.Debugf("keepalive: %d\n", ka.TTL)
}(ch)
log.Debug("register finished")
idx = i
registered = true
@ -207,7 +201,7 @@ func (e *EtcdClient) registerPserverEtcd(ctx context.Context, port int) (int, er
// GetKey gets the value by the specified key
func (e *EtcdClient) GetKey(key string, timeout time.Duration) ([]byte, error) {
ctx, cancel := context.WithTimeout(context.Background(), timeout)
resp, err := e.etcdClient.Get(ctx, key)
resp, err := e.client.Get(ctx, key)
cancel()
if err != nil {
return []byte{}, err
@ -223,7 +217,27 @@ func (e *EtcdClient) GetKey(key string, timeout time.Duration) ([]byte, error) {
// PutKey put into etcd with value by key specified
func (e *EtcdClient) PutKey(key string, value []byte, timeout time.Duration) error {
ctx, cancel := context.WithTimeout(context.Background(), timeout)
_, err := e.etcdClient.Put(ctx, key, string(value))
_, err := e.client.Put(ctx, key, string(value), clientv3.WithLease(e.sess.Lease()))
cancel()
return err
}
// Shutdown shuts down the etcd client gracefully.
func (e *EtcdClient) Shutdown() error {
var err error
if e.sess != nil {
err = e.sess.Close()
}
if e.client != nil {
newErr := e.client.Close()
if newErr != nil {
if err != nil {
log.Errorln(newErr)
} else {
err = newErr
}
}
}
return err
}

@ -269,8 +269,7 @@ void hl_sequence2batch_copy_padding(real* batch,
int blockDimY = CUDA_BLOCK_SIZE / blockDimX;
dim3 threads(blockDimX, blockDimY);
int gridDimX = (maxSequenceLength * blockDimX + CUDA_BLOCK_SIZE - 1) /
CUDA_BLOCK_SIZE;
int gridDimX = (maxSequenceLength + blockDimY - 1) / blockDimY;
int gridDimY = numSequences;
dim3 grid(gridDimX, gridDimY);

@ -83,56 +83,38 @@ inline void Tensor::ShareDataWith(const Tensor& src) {
template <typename T>
inline void Tensor::CopyFrom(const Tensor& src,
const platform::CPUDeviceContext& ctx) {
const platform::Place& dst_place) {
src.check_memory_size<T>();
Resize(src.dims());
auto src_place = src.holder_->place();
auto src_ptr = static_cast<const void*>(src.data<T>());
auto dst_place = ctx.GetPlace();
auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
auto size = product(src.dims_) * sizeof(T);
if (platform::is_cpu_place(src_place)) {
if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
boost::get<platform::CPUPlace>(src_place), src_ptr, size);
}
#ifndef PADDLE_ONLY_CPU
else if (platform::is_gpu_place(src_place)) {
else if (platform::is_gpu_place(src_place) &&
platform::is_cpu_place(dst_place)) {
memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
boost::get<platform::GPUPlace>(src_place), src_ptr, size, 0);
}
#endif
}
#ifndef PADDLE_ONLY_CPU
template <typename T>
inline void Tensor::CopyFrom(const Tensor& src,
const platform::CUDADeviceContext& ctx) {
src.check_memory_size<T>();
Resize(src.dims());
auto src_place = src.holder_->place();
auto src_ptr = static_cast<const void*>(src.data<T>());
auto dst_place = ctx.GetPlace();
auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
auto size = product(src.dims_) * sizeof(T);
if (platform::is_cpu_place(src_place)) {
} else if (platform::is_cpu_place(src_place) &&
platform::is_gpu_place(dst_place)) {
memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr,
boost::get<platform::CPUPlace>(src_place), src_ptr, size,
ctx.stream());
} else if (platform::is_gpu_place(src_place)) {
boost::get<platform::CPUPlace>(src_place), src_ptr, size, 0);
} else if (platform::is_gpu_place(src_place) &&
platform::is_gpu_place(dst_place)) {
memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr,
boost::get<platform::GPUPlace>(src_place), src_ptr, size,
ctx.stream());
boost::get<platform::GPUPlace>(src_place), src_ptr, size, 0);
}
}
#endif
}
template <typename T>
inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {

@ -11,8 +11,7 @@ static int run_cnt = 0;
class TestOp : public OperatorBase {
public:
void InferShape(
const std::shared_ptr<framework::Scope>& scope) const override {
void InferShape(const std::shared_ptr<Scope>& scope) const override {
++infer_shape_cnt;
}
void Run(const std::shared_ptr<framework::Scope>& scope,

@ -20,7 +20,7 @@ namespace paddle {
namespace framework {
template <>
Eigen::DefaultDevice* KernelContext::GetEigenDevice<
Eigen::DefaultDevice* ExecutionContext::GetEigenDevice<
platform::CPUPlace, Eigen::DefaultDevice>() const {
return device_context_.get_eigen_device<Eigen::DefaultDevice>();
}
@ -28,7 +28,7 @@ Eigen::DefaultDevice* KernelContext::GetEigenDevice<
#ifndef PADDLE_ONLY_CPU
template <>
Eigen::GpuDevice*
KernelContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
return device_context_.get_eigen_device<Eigen::GpuDevice>();
}
#endif

File diff suppressed because it is too large Load Diff

@ -24,7 +24,8 @@ static int op_run_num = 0;
class OpWithoutKernelTest : public OperatorBase {
public:
void Init() override { x = 1; }
void InferShape(const std::shared_ptr<Scope>& scope) const override {}
void InferShape(
const std::shared_ptr<framework::Scope>& scope) const override {}
void Run(const std::shared_ptr<Scope>& scope,
const platform::DeviceContext& dev_ctx) const override {
op_run_num++;
@ -73,6 +74,7 @@ TEST(OperatorBase, all) {
auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
scope->CreateVariable("OUT1");
ASSERT_EQ(paddle::framework::op_run_num, 0);
op->InferShape(scope);
op->Run(scope, device_context);
ASSERT_EQ(paddle::framework::op_run_num, 1);
}
@ -97,14 +99,13 @@ static int cpu_kernel_run_num = 0;
class OpWithKernelTest : public OperatorWithKernel {
protected:
void InferShape(const std::vector<const Tensor*>& inputs,
const std::vector<Tensor*>& outputs) const override {}
void InferShape(const framework::InferShapeContext& ctx) const override {}
};
template <typename T1, typename T2>
class CPUKernelTest : public OpKernel {
public:
void Compute(const KernelContext& ctx) const {
void Compute(const ExecutionContext& ctx) const {
std::cout << "this is cpu kernel" << std::endl;
std::cout << ctx.op_.DebugString() << std::endl;
cpu_kernel_run_num++;
@ -117,7 +118,8 @@ class CPUKernelTest : public OpKernel {
class OperatorMultiInputsTest : public OperatorBase {
public:
void Init() override { x = 1; }
void InferShape(const std::shared_ptr<Scope>& scope) const override {}
void InferShape(
const std::shared_ptr<framework::Scope>& scope) const override {}
void Run(const std::shared_ptr<Scope>& scope,
const platform::DeviceContext& dev_ctx) const override {
ASSERT_EQ(scope->GetVariable(inputs_[0]), nullptr);
@ -149,13 +151,31 @@ class OpKernelTestMultiInputsProtoAndCheckerMaker
class CPUKernalMultiInputsTest : public OpKernel {
public:
void Compute(const KernelContext& ctx) const {
void Compute(const ExecutionContext& ctx) const {
auto xs = ctx.op_.Inputs("xs");
ASSERT_EQ(xs.size(), 3UL);
ASSERT_EQ(xs[0], "x0");
ASSERT_EQ(xs[1], "x1");
ASSERT_EQ(xs[2], "x2");
auto inVar0 = ctx.MultiInputVar("xs");
ASSERT_EQ(inVar0.size(), 3);
auto intVar1 = ctx.InputVar("k");
ASSERT_NE(intVar1, nullptr);
auto outVar0 = ctx.MultiOutputVar("ys");
ASSERT_EQ(outVar0.size(), 2);
auto inTensor0 = ctx.MultiInput<Tensor>("xs");
ASSERT_EQ(inTensor0.size(), 3);
auto intTensor1 = ctx.Input<Tensor>("k");
ASSERT_NE(intTensor1, nullptr);
auto outTensor0 = ctx.MultiOutput<Tensor>("ys");
ASSERT_EQ(outTensor0.size(), 2);
auto k = ctx.op_.Input("k");
ASSERT_EQ(k, "k0");
@ -233,6 +253,12 @@ TEST(OpKernel, multi_inputs) {
paddle::platform::CPUDeviceContext cpu_device_context;
auto scope = std::make_shared<Scope>();
scope->CreateVariable("x0")->GetMutable<Tensor>();
scope->CreateVariable("x1")->GetMutable<Tensor>();
scope->CreateVariable("x2")->GetMutable<Tensor>();
scope->CreateVariable("k0")->GetMutable<Tensor>();
scope->CreateVariable("y0")->GetMutable<Tensor>();
scope->CreateVariable("y1")->GetMutable<Tensor>();
auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
op->Run(scope, cpu_device_context);

@ -94,14 +94,7 @@ class Tensor {
* @note CopyFrom supports CPU <-> GPU, GPU <-> GPU.
*/
template <typename T>
inline void CopyFrom(const Tensor& src,
const platform::CPUDeviceContext& ctx);
#ifndef PADDLE_ONLY_CPU
template <typename T>
inline void CopyFrom(const Tensor& src,
const platform::CUDADeviceContext& ctx);
#endif
inline void CopyFrom(const Tensor& src, const platform::Place& dst_place);
/**
* @brief Return the slice of the tensor.
@ -129,13 +122,16 @@ class Tensor {
virtual platform::Place place() const = 0;
};
template <typename T, typename PlaceType>
template <typename T, typename Place>
struct PlaceholderImpl : public Placeholder {
PlaceholderImpl(PlaceType place, size_t size)
PlaceholderImpl(Place place, size_t size)
: ptr_(static_cast<T*>(memory::Alloc(place, size)),
memory::PODDeleter<T, PlaceType>(place)),
memory::PODDeleter<T, Place>(place)),
place_(place),
size_(size) {}
size_(size) {
PADDLE_ENFORCE(ptr_ != nullptr, "Insufficient %s memory to allocation.",
is_cpu_place(place_) ? "CPU" : "GPU");
}
virtual size_t size() const { return size_; }
virtual platform::Place place() const { return place_; }
@ -143,7 +139,7 @@ class Tensor {
virtual std::type_index type() const { return std::type_index(typeid(T)); }
/*! the pointer of memory block. */
std::unique_ptr<T, memory::PODDeleter<T, PlaceType>> ptr_;
std::unique_ptr<T, memory::PODDeleter<T, Place>> ptr_;
/*! the place of memory block. */
platform::Place place_;

@ -198,8 +198,8 @@ TEST(Tensor, CopyFrom) {
int arr[9] = {1, 2, 3, 4, 5, 6, 7, 8, 9};
memcpy(src_ptr, arr, 9 * sizeof(int));
auto* cpu_ctx = new paddle::platform::CPUDeviceContext();
dst_tensor.CopyFrom<int>(src_tensor, *cpu_ctx);
auto cpu_place = new paddle::platform::CPUPlace();
dst_tensor.CopyFrom<int>(src_tensor, *cpu_place);
const int* dst_ptr = dst_tensor.data<int>();
ASSERT_NE(src_ptr, dst_ptr);
@ -208,7 +208,7 @@ TEST(Tensor, CopyFrom) {
}
Tensor slice_tensor = src_tensor.Slice<int>(1, 2);
dst_tensor.CopyFrom<int>(slice_tensor, *cpu_ctx);
dst_tensor.CopyFrom<int>(slice_tensor, *cpu_place);
const int* slice_ptr = slice_tensor.data<int>();
dst_ptr = dst_tensor.data<int>();
ASSERT_NE(dst_ptr, slice_ptr);
@ -228,12 +228,12 @@ TEST(Tensor, CopyFrom) {
memcpy(src_ptr, arr, 9 * sizeof(int));
// CPU Tensor to GPU Tensor
auto gpu_ctx = new paddle::platform::CUDADeviceContext(0);
gpu_tensor.CopyFrom<int>(src_tensor, *gpu_ctx);
auto gpu_place = new paddle::platform::GPUPlace(0);
gpu_tensor.CopyFrom<int>(src_tensor, *gpu_place);
// GPU Tensor to CPU Tensor
auto cpu_ctx = new paddle::platform::CPUDeviceContext();
dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_ctx);
auto cpu_place = new paddle::platform::CPUPlace();
dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place);
// Compare Tensors
const int* dst_ptr = dst_tensor.data<int>();
@ -245,10 +245,10 @@ TEST(Tensor, CopyFrom) {
Tensor slice_tensor = src_tensor.Slice<int>(1, 2);
// CPU Slice Tensor to GPU Tensor
gpu_tensor.CopyFrom<int>(slice_tensor, *gpu_ctx);
gpu_tensor.CopyFrom<int>(slice_tensor, *gpu_place);
// GPU Tensor to CPU Tensor
dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_ctx);
dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place);
// Compare Slice Tensors
const int* slice_ptr = slice_tensor.data<int>();

@ -1141,4 +1141,64 @@ TEST(CpuMatrix, copyFrom) {
TensorCheckEqual(cpu, copy);
}
void testBatch2seqPadding(int batchSize, int inputDim) {
MatrixPtr cpuInput = std::make_shared<CpuMatrix>(batchSize, inputDim);
MatrixPtr gpuInput = std::make_shared<GpuMatrix>(batchSize, inputDim);
cpuInput->randomizeUniform();
gpuInput->copyFrom(*cpuInput);
IVectorPtr cpuSequence;
generateSequenceStartPositions(batchSize, cpuSequence);
IVectorPtr gpuSequence = IVector::create(cpuSequence->getSize(), true);
gpuSequence->copyFrom(*cpuSequence);
size_t numSeq = cpuSequence->getSize() - 1;
size_t maxSeqLen = *std::max_element(cpuSequence->getData(),
cpuSequence->getData() + numSeq);
MatrixPtr cBatch = std::make_shared<CpuMatrix>(numSeq * maxSeqLen, inputDim);
MatrixPtr gBatch = std::make_shared<GpuMatrix>(numSeq * maxSeqLen, inputDim);
MatrixPtr cCheck = std::make_shared<CpuMatrix>(numSeq * maxSeqLen, inputDim);
hl_sequence2batch_copy_padding(gBatch->getData(),
gpuInput->getData(),
cpuSequence->getData(),
inputDim,
maxSeqLen,
numSeq,
false,
true);
cCheck->copyFrom(*gBatch);
int* seqStart = cpuSequence->getData();
float* batchData = cBatch->getData();
float* seqData = cpuInput->getData();
for (size_t i = 0; i < maxSeqLen; i++) {
for (size_t j = 0; j < numSeq; j++) {
size_t sequenceStart = seqStart[j];
size_t sequenceLength = seqStart[j + 1] - seqStart[j];
if (i < sequenceLength) {
memcpy(batchData + (i * numSeq + j) * inputDim,
seqData + (sequenceStart + i) * inputDim,
inputDim * sizeof(real));
} else {
memset(batchData + (i * numSeq + j) * inputDim,
0,
inputDim * sizeof(real));
}
}
}
TensorCheckErr(*cBatch, *cCheck);
}
TEST(Matrix, warpCTC) {
for (auto batchSize : {51, 526, 2884}) {
for (auto inputDim : {32, 512, 2026}) {
VLOG(3) << " batchSize=" << batchSize << " inputDim=" << inputDim;
testBatch2seqPadding(batchSize, inputDim);
}
}
}
#endif

@ -35,7 +35,7 @@ void Copy<platform::CPUPlace, platform::GPUPlace>(platform::CPUPlace dst_place,
platform::GPUPlace src_place,
const void* src, size_t num,
cudaStream_t stream) {
platform::GPUPlaceGuard g(src_place.device);
platform::SetDeviceId(src_place.device);
platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyDeviceToHost, stream);
}
@ -45,7 +45,7 @@ void Copy<platform::GPUPlace, platform::CPUPlace>(platform::GPUPlace dst_place,
platform::CPUPlace src_place,
const void* src, size_t num,
cudaStream_t stream) {
platform::GPUPlaceGuard g(dst_place.device);
platform::SetDeviceId(dst_place.device);
platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyHostToDevice, stream);
}
@ -56,7 +56,7 @@ void Copy<platform::GPUPlace, platform::GPUPlace>(platform::GPUPlace dst_place,
const void* src, size_t num,
cudaStream_t stream) {
if (dst_place == src_place) {
platform::GPUPlaceGuard g(src_place.device);
platform::SetDeviceId(src_place.device);
platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyDeviceToDevice, stream);
} else {
platform::GpuMemcpyPeer(dst, dst_place.device, src, src_place.device, num,

@ -20,13 +20,39 @@ limitations under the License. */
namespace paddle {
namespace memory {
/**
* \brief Copy memory from one place to another place.
*
* \param[in] DstPlace Destination allocation place (CPU).
* \param[in] dst Destination memory address.
* \param[in] SrcPlace Source allocation place (CPU).
* \param[in] src Source memory address.
* \param[in] num memory size in bytes to copy.
*
*/
template <typename DstPlace, typename SrcPlace>
void Copy(DstPlace, void* dst, SrcPlace, const void* src, size_t num);
#ifndef PADDLE_ONLY_CPU
/**
* \brief Copy memory from one place to another place.
*
* \param[in] DstPlace Destination allocation place (CPU or GPU).
* \param[in] dst Destination memory address.
* \param[in] SrcPlace Source allocation place (CPU or GPU).
* \param[in] src Source memory address.
* \param[in] num memory size in bytes to copy.
* \param[in] stream CUDA stream.
*
* \note For GPU memory copy, CUDA stream need to be specified
* for asynchronously memory copy.
*
*/
template <typename DstPlace, typename SrcPlace>
void Copy(DstPlace, void* dst, SrcPlace, const void* src, size_t num,
cudaStream_t stream);
#endif // PADDLE_ONLY_CPU
} // namespace memory

@ -60,6 +60,7 @@ detail::BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
platform::GpuMaxChunkSize());
}
}
platform::SetDeviceId(gpu_id);
return as[gpu_id];
}

@ -20,15 +20,49 @@ limitations under the License. */
namespace paddle {
namespace memory {
/**
* \brief Allocate memory block in one place.
*
* \param[in] place Allocation place (CPU or GPU).
* \param[in] size Allocation size.
*
* \return Allocated memory block address.
*
* \note If return nullptr, it indicates memory allocation failed
* because insufficient memory in current system. When Alloc
* function is invoked, you must check the returned memory
* address is valid or not.
*/
template <typename Place>
void* Alloc(Place, size_t);
void* Alloc(Place place, size_t size);
/**
* \brief Free memory block in one place.
*
* \param[in] place Allocation place (CPU or GPU).
* \param[in] ptr Memory block address to free.
*
*/
template <typename Place>
void Free(Place, void*);
void Free(Place place, void* ptr);
/**
* \brief Total size of used memory in one place.
*
* \param[in] place Allocation place (CPU or GPU).
*
*/
template <typename Place>
size_t Used(Place);
size_t Used(Place place);
/**
* \brief Free memory block in one place.
*
* \note In some cases, custom deleter is used to
* deallocate the memory automatically for
* std::unique_ptr<T> in tensor.h.
*
*/
template <typename T, typename Place>
class PODDeleter {
static_assert(std::is_pod<T>::value, "T must be POD");

Some files were not shown because too many files have changed in this diff Show More

Loading…
Cancel
Save