portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2026-08-31 14:59:35 +01:00
Code Releases Activity

Better ROCTX tracing

Browse Source
This commit is contained in:
Peter Boyle
2026-08-20 11:46:33 -04:00
parent 3ed05dbbb1
commit d62fc91996
7 changed files with 25 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Grid/perfmon/Tracing.h
-2
View File
@@ -25,11 +25,9 @@ class GridTracer {
public:
GridTracer(const char* name) {
roctxRangePushA(name);
std::cout << "roctxRangePush "<<name<<std::endl;
}
~GridTracer() {
roctxRangePop();
std::cout << "roctxRangePop "<<std::endl;
}
};
inline void tracePush(const char *name) { roctxRangePushA(name); }
Grid/stencil/Stencil.h
+4
View File
@@ -314,6 +314,7 @@ public:
return accessor;
}
int traceID;
int face_table_computed;
// int partialDirichlet;
int fullDirichlet;
@@ -537,6 +538,7 @@ public:
_grid->StencilBarrier();
#endif
}
traceID = traceStart("Stencil::CommunicateBegin");
for(int i=0;i<Packets.size();i++){
// std::cout << "Communicate prepare "<<i<<std::endl;
@@ -588,6 +590,8 @@ public:
// _grid->Barrier();
_grid->StencilSendToRecvFromComplete(MpiReqs,0); // MPI is done
// if ( this->partialDirichlet ) DslashLogPartial();
traceStop(traceID);
if ( this->fullDirichlet ) DslashLogDirichlet();
else DslashLogFull();
// acceleratorCopySynchronise();// is in the StencilSendToRecvFromComplete
Grid/threads/Accelerator.h
+4 -5
View File
@@ -251,7 +251,7 @@ inline void acceleratorFreeDevice(void *ptr){ cudaFree(ptr);};
inline void acceleratorFreeHost(void *ptr){ cudaFree(ptr);};
inline void acceleratorCopyToDevice(const void *from,void *to,size_t bytes) { cudaMemcpy(to,from,bytes, cudaMemcpyHostToDevice);}
inline void acceleratorCopyFromDevice(const void *from,void *to,size_t bytes){ cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToHost);}
inline void acceleratorMemSet(void *base,int value,size_t bytes) { cudaMemset(base,value,bytes);}
inline void acceleratorMemSet(void *base,int value,size_t bytes) { cudaMemset(base,value,bytes); cudaDeviceSynchronize(); }
inline acceleratorEvent_t acceleratorCopyToDeviceAsynch(void *from, void *to, size_t bytes, cudaStream_t stream = copyStream) {
acceleratorCopyToDevice(from,to,bytes);
return 0;
@@ -525,10 +525,9 @@ inline void *acceleratorAllocDevice(size_t bytes)
inline void acceleratorFreeHost(void *ptr){ auto discard=hipFree(ptr);};
inline void acceleratorFreeShared(void *ptr){ auto discard=hipFree(ptr);};
inline void acceleratorFreeDevice(void *ptr){ auto discard=hipFree(ptr);};
inline void acceleratorCopyToDevice(const void *from,void *to,size_t bytes) { auto discard=hipMemcpy(to,from,bytes, hipMemcpyHostToDevice);}
inline void acceleratorCopyFromDevice(const void *from,void *to,size_t bytes){ auto discard=hipMemcpy(to,from,bytes, hipMemcpyDeviceToHost);}
inline void acceleratorMemSet(void *base,int value,size_t bytes) { auto discard=hipMemset(base,value,bytes);}
inline void acceleratorCopyToDevice(const void *from,void *to,size_t bytes) { auto discard=hipMemcpy(to,from,bytes, hipMemcpyHostToDevice); }
inline void acceleratorCopyFromDevice(const void *from,void *to,size_t bytes){ auto discard=hipMemcpy(to,from,bytes, hipMemcpyDeviceToHost); }
inline void acceleratorMemSet(void *base,int value,size_t bytes) { auto discard=hipMemset(base,value,bytes); discard = hipDeviceSynchronize(); }
typedef int acceleratorEvent_t;