[topo-expl] update header file location. (#1769)
Cette révision appartient à :
révisé par
GitHub
Parent
16101e654f
révision
35024ca1cb
@@ -6,7 +6,7 @@ endif
|
||||
HIPCC = $(HIP_PATH)/bin/hipcc
|
||||
|
||||
EXE = topo_expl
|
||||
CXXFLAGS = -g -ffunction-sections -fdata-sections -Wl,--gc-sections -fgpu-rdc -Iinclude -Ihipify_rccl/include -Ihipify_rccl/device/include -Ihipify_rccl/graph -I/opt/rocm/include/ -DTOPO_EXPL -DENABLE_TRACE -DENABLE_LL128 -DNVTX_NO_IMPL -DRCCL_EXPOSE_STATIC -lpthread
|
||||
CXXFLAGS = -g -ffunction-sections -fdata-sections -Wl,--gc-sections -fgpu-rdc -Iinclude -Ihipify_rccl/include -Ihipify_rccl/include/plugin -Ihipify_rccl/device/include -Ihipify_rccl/graph -I/opt/rocm/include/ -DTOPO_EXPL -DENABLE_TRACE -DENABLE_LL128 -DNVTX_NO_IMPL -DRCCL_EXPOSE_STATIC -lpthread
|
||||
|
||||
files = $(EXE).cpp model.cpp utils.cpp hipify_rccl/graph/topo.cc hipify_rccl/graph/rings.cc hipify_rccl/graph/paths.cc hipify_rccl/graph/trees.cc ../../src/misc/param.cc \
|
||||
hipify_rccl/graph/search.cc hipify_rccl/graph/connect.cc hipify_rccl/graph/tuning.cc hipify_rccl/graph/xml.cc ../../src/misc/nvmlwrap_stub.cc hipify_rccl/graph/rome_models.cc hipify_rccl/graph/archinfo.cc \
|
||||
@@ -29,6 +29,7 @@ hipify:
|
||||
cp -a ../../src/collectives.cc hipify_rccl/
|
||||
cp -a ../../src/misc/archinfo.cc hipify_rccl/graph/
|
||||
hipify-perl -inplace -quiet-warnings hipify_rccl/include/*.h
|
||||
hipify-perl -inplace -quiet-warnings hipify_rccl/include/plugin/*.h
|
||||
hipify-perl -inplace -quiet-warnings hipify_rccl/device/include/*.h
|
||||
sed -i "s/template<typename T, typename RedOp>/template<typename T, typename RedOp, int COLL_UNROLL>/g" "hipify_rccl/device/include/common.h"
|
||||
sed -i "s/\\(struct RunWorkBatch<ncclFunc[^>]*\\)>*/\\1, COLL_UNROLL>/" "hipify_rccl/device/include/common.h"
|
||||
@@ -38,4 +39,4 @@ hipify:
|
||||
|
||||
clean:
|
||||
rm -rf hipify_rccl
|
||||
rm -f *.o $(EXE)
|
||||
rm -f *.o $(EXE)
|
||||
|
||||
@@ -191,7 +191,7 @@ struct setupReq {
|
||||
int remoteRank;
|
||||
int shared;
|
||||
int netDev;
|
||||
int useGdr;
|
||||
enum ncclTopoGdrMode useGdr;
|
||||
int channelId;
|
||||
int connIndex;
|
||||
};
|
||||
@@ -263,7 +263,8 @@ ncclResult_t collNetCanConnect(int* ret, struct ncclComm* comm, struct ncclTopo
|
||||
}
|
||||
|
||||
ncclResult_t collNetSendSetup(struct ncclComm* comm, struct ncclTopoGraph* graph, struct ncclPeerInfo* myInfo, struct ncclPeerInfo* peerInfo, struct ncclConnect* connectInfo, struct ncclConnector* send, int channelId, int connIndex) {
|
||||
int netDev, useGdr = 0, proxy;
|
||||
int netDev, proxy;
|
||||
enum ncclTopoGdrMode useGdr = ncclTopoGdrModeDisable;
|
||||
|
||||
NCCLCHECK(ncclTopoGetNetDev(comm, myInfo->rank, graph, channelId, peerInfo->rank, NULL, &netDev, &proxy));
|
||||
NCCLCHECK(ncclTopoCheckGdr(comm->topo, myInfo->rank, netDev, 1, &useGdr));
|
||||
@@ -273,7 +274,8 @@ ncclResult_t collNetSendSetup(struct ncclComm* comm, struct ncclTopoGraph* graph
|
||||
}
|
||||
|
||||
ncclResult_t collNetRecvSetup(struct ncclComm* comm, struct ncclTopoGraph* graph, struct ncclPeerInfo* myInfo, struct ncclPeerInfo* peerInfo, struct ncclConnect* connectInfo, struct ncclConnector* recv, int channelId, int connIndex) {
|
||||
int netDev, useGdr = 0, proxy;
|
||||
int netDev, proxy;
|
||||
enum ncclTopoGdrMode useGdr = ncclTopoGdrModeDisable;
|
||||
|
||||
NCCLCHECK(ncclTopoGetNetDev(comm, myInfo->rank, graph, channelId, peerInfo->rank, NULL, &netDev, &proxy));
|
||||
NCCLCHECK(ncclTopoCheckGdr(comm->topo, myInfo->rank, netDev, 0, &useGdr));
|
||||
|
||||
Référencer dans un nouveau ticket
Bloquer un utilisateur