bda/gpu-mpi-tests
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

add mpi wtime counters, fix make clean

Browse Source
This commit is contained in:
Bryce Allen
2020-08-07 13:05:34 -04:00
parent 3dd6045f2e
commit 3ebd09725e
2 changed files with 19 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
Makefile
View File

@@ -21,7 +21,7 @@ mpienv: mpienv.f90
.PHONY: clean
clean:
rm -rf daxpy mpi_daxpy
rm -rf daxpy mpi_daxpy daxpy_nvtx mpi_daxpy_nvtx
.PHONY: force
force: clean all

19
mpi_daxpy_nvtx.cc
View File

@@ -66,12 +66,19 @@ void set_rank_device(int n_ranks, int rank) {
int main(int argc, char **argv) {
int n = 1024;
int n = 32*1024*1024;
int world_size, world_rank;
double a = 2.0;
double sum = 0.0;
double start_time = 0.0;
double end_time = 0.0;
double k_start_time = 0.0;
double k_end_time = 0.0;
double g_start_time = 0.0;
double g_end_time = 0.0;
//double *x, *y, *d_x, *d_y;
double *m_x, *m_y;
@@ -113,6 +120,7 @@ int main(int argc, char **argv) {
//CHECK("setDevice", cudaSetDevice(0));
cudaProfilerStart();
start_time = MPI_Wtime();
CHECK( "cublas", cublasCreate(&handle) );
@@ -155,12 +163,14 @@ int main(int argc, char **argv) {
MEMINFO("m_x", m_x, sizeof(m_x));
MEMINFO("m_y", m_y, sizeof(m_y));
k_start_time = MPI_Wtime();
nvtxRangePushA("cublasDaxpy");
CHECK("daxpy",
cublasDaxpy(handle, n, &a, m_x, 1, m_y, 1) );
CHECK("daxpy sync", cudaDeviceSynchronize());
nvtxRangePop();
k_end_time = MPI_Wtime();
/*
CHECK("y = d_y",
@@ -182,6 +192,7 @@ int main(int argc, char **argv) {
nvtxRangePop();
printf("%d/%d SUM = %f\n", world_rank, world_size, sum);
g_start_time = MPI_Wtime();
nvtxRangePushA("allGather");
nvtxRangePushA("x");
MPI_Allgather(m_x, n, MPI_DOUBLE, m_allx, n, MPI_DOUBLE, MPI_COMM_WORLD);
@@ -190,6 +201,7 @@ int main(int argc, char **argv) {
MPI_Allgather(m_y, n, MPI_DOUBLE, m_ally, n, MPI_DOUBLE, MPI_COMM_WORLD);
nvtxRangePop();
nvtxRangePop();
g_end_time = MPI_Wtime();
sum = 0.0;
nvtxRangePushA("allSum");
@@ -211,10 +223,15 @@ int main(int argc, char **argv) {
nvtxRangePop();
end_time = MPI_Wtime();
cudaProfilerStop();
cublasDestroy(handle);
MPI_Finalize();
printf("total time: %0.3f\n", end_time-start_time);
printf("kernel time: %0.3f\n", k_end_time-k_start_time);
printf("gather time: %0.3f\n", g_end_time-g_start_time);
return EXIT_SUCCESS;
}