gallo/atrip
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Improve cuda m4

Browse Source
This commit is contained in:
Gallo Alejandro 2022-08-11 13:55:52 +02:00
parent f06cd7f562
commit 4101c89907
2 changed files with 70 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
bench/test_main.cxx
View File

@ -17,7 +17,7 @@ int main(int argc, char** argv) {
MPI_Init(&argc, &argv); MPI_Init(&argc, &argv);
size_t checkpoint_it; size_t checkpoint_it;
int no(10), nv(10), itMod(-1), percentageMod(10); int no(10), nv(100), itMod(-1), percentageMod(10);
float checkpoint_percentage; float checkpoint_percentage;
bool bool
nochrono(false), barrier(false), rankRoundRobin(false), nochrono(false), barrier(false), rankRoundRobin(false),

70
etc/m4/atrip_cublas.m4
View File

@ -21,7 +21,7 @@ ac_save_LDFLAGS="$LDFLAGS"
CXXFLAGS="${CXXFLAGS} $CUDA_CXXFLAGS" CXXFLAGS="${CXXFLAGS} $CUDA_CXXFLAGS"
LDFLAGS="${LDFLAGS} $CUDA_LDFLAGS" LDFLAGS="${LDFLAGS} $CUDA_LDFLAGS"
AC_MSG_CHECKING([that cublas works with $CXX]) AC_MSG_CHECKING([that cublas works with the CXX compiler])
AC_COMPILE_IFELSE([AC_LANG_SOURCE([_ATRIP_CUBLAS_SOURCE])], AC_COMPILE_IFELSE([AC_LANG_SOURCE([_ATRIP_CUBLAS_SOURCE])],
[ [
@ -33,6 +33,16 @@ AC_COMPILE_IFELSE([AC_LANG_SOURCE([_ATRIP_CUBLAS_SOURCE])],
AC_MSG_ERROR([Does not work!]) AC_MSG_ERROR([Does not work!])
]) ])
AC_MSG_CHECKING([that the detected graphic cards have memory available])
AC_RUN_IFELSE([AC_LANG_SOURCE([_ATRIP_CUDA_MEMORY_OF_DEVICES])],
[
atrip_success=yes
],
[
atrip_success=no
AC_MSG_ERROR([An available device reports zero memory available!])
])
CXX="$ac_save_CXX" CXX="$ac_save_CXX"
CXXFLAGS="$ac_save_CXXFLAGS" CXXFLAGS="$ac_save_CXXFLAGS"
LDFLAGS="$ac_save_LDFLAGS" LDFLAGS="$ac_save_LDFLAGS"
@ -79,3 +89,61 @@ int main() {
return 0; return 0;
} }
]]) ]])
m4_define([_ATRIP_CUDA_MEMORY_OF_DEVICES], [[
#include <mpi.h>
#include <iostream>
#include <cassert>
#include <string.h>
#include <cuda.h>
int main() {
MPI_Init(NULL, NULL);
int rank, np;
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &np);
int ngcards;
cuDeviceGetCount(&ngcards);
for (size_t rank = 0; rank < ngcards; rank++) {
CUcontext ctx;
CUdevice dev;
CUdevprop_st prop;
size_t _free, total, total2;
char *name = (char*)malloc(256);
cuDeviceGet(&dev, rank);
cuCtxCreate(&ctx, 0, dev);
cuCtxSetCurrent(ctx);
cuDeviceGetProperties(&prop, dev);
cuMemGetInfo(&_free, &total);
cuDeviceGetName(name, 256, dev);
cuDeviceTotalMem(&total2, dev);
printf("\n"
"CUDA CARD RANK %d\n"
"=================\n"
"\tname: %s\n"
"\tShared Mem Per Block (KB): %f\n"
"\tFree/Total mem (GB): %f/%f\n"
"\ttotal2 mem (GB): %f\n"
"\n",
dev,
name,
prop.sharedMemPerBlock / 1024.0,
_free / 1024.0 / 1024.0 / 1024.0 ,
total / 1024.0 / 1024.0 / 1024.0 ,
total2 / 1024.0 / 1024.0 / 1024.0
);
if (_free == 0 || total == 0 || total2 == 0)
return 1;
}
return 0;
}
]])