Skip to content
Snippets Groups Projects
Commit ea94cf82 authored by Giovanni Lacopo's avatar Giovanni Lacopo
Browse files

Memory bug fixing

parent 011d30c6
No related branches found
No related tags found
No related merge requests found
Show whitespace changes
Inline Side-by-side
gridding_cpu.c
+ 21
− 30
View file @ ea94cf82
...@@ -84,17 +84,15 @@ void gridding_data() ...@@ -84,17 +84,15 @@ void gridding_data()
// find the largest value in histo_send[] // find the largest value in histo_send[]
// //
uint Nsec = histo_send[0];
for (uint isector = 1; isector < nsectors; isector++)
Nsec = ( Nsec < histo_send[isector] ? histo_send[isector] : Nsec );
for (uint isector = 0; isector < nsectors; isector++)
{
double start = CPU_TIME_wt;
uint Nsec = histo_send[isector];
uint Nweightss = Nsec*metaData.polarisations; uint Nweightss = Nsec*metaData.polarisations;
uint Nvissec = Nweightss*metaData.freq_per_chan; uint Nvissec = Nweightss*metaData.freq_per_chan;
// allocate sector arrays
// note: we use the largest allocation among all sectors
//
unsigned long long int mem_size = (Nsec*3)*sizeof(double_t) + (Nvissec*2+Nweightss)*sizeof(float_t);
double_t *memory = (double*) malloc ( (Nsec*3)*sizeof(double_t) + double_t *memory = (double*) malloc ( (Nsec*3)*sizeof(double_t) +
(Nvissec*2+Nweightss)*sizeof(float_t) ); (Nvissec*2+Nweightss)*sizeof(float_t) );
...@@ -109,12 +107,6 @@ void gridding_data() ...@@ -109,12 +107,6 @@ void gridding_data()
float_t *visimgs = (float_t*)visreals + Nvissec; float_t *visimgs = (float_t*)visreals + Nvissec;
for (uint isector = 0; isector < nsectors; isector++)
{
double start = CPU_TIME_wt;
memset( memory, 0, mem_size );
// select data for this sector // select data for this sector
uint icount = 0; uint icount = 0;
...@@ -183,14 +175,14 @@ void gridding_data() ...@@ -183,14 +175,14 @@ void gridding_data()
printf("Processing sector %ld\n",isector); printf("Processing sector %ld\n",isector);
#endif #endif
start = CPU_TIME_wt;
double *stacking_target_array; double *stacking_target_array;
if ( size > 1 ) if ( size > 1 )
stacking_target_array = gridss; stacking_target_array = gridss;
else else
stacking_target_array = grid; stacking_target_array = grid;
start = CPU_TIME_wt;
//We have to call different GPUs per MPI task!!! [GL] //We have to call different GPUs per MPI task!!! [GL]
wstack(param.num_w_planes, wstack(param.num_w_planes,
Nsec, Nsec,
...@@ -259,9 +251,8 @@ void gridding_data() ...@@ -259,9 +251,8 @@ void gridding_data()
memset ( gridss, 0, 2*param.num_w_planes*xaxis*yaxis * sizeof(double) ); memset ( gridss, 0, 2*param.num_w_planes*xaxis*yaxis * sizeof(double) );
} }
}
free(memory); free(memory);
}
if ( size > 1 ) if ( size > 1 )
{ {
......
gridding_nccl.cpp
+ 24
− 31
View file @ ea94cf82
...@@ -68,30 +68,6 @@ void gridding_data(){ ...@@ -68,30 +68,6 @@ void gridding_data(){
// find the largest value in histo_send[] // find the largest value in histo_send[]
// //
uint Nsec = histo_send[0];
for (uint isector = 1; isector < nsectors; isector++)
Nsec = ( Nsec < histo_send[isector] ? histo_send[isector] : Nsec );
uint Nweightss = Nsec*metaData.polarisations;
uint Nvissec = Nweightss*metaData.freq_per_chan;
// allocate sector arrays
// note: we use the largest allocation among all sectors
//
unsigned long long int mem_size = (Nsec*3)*sizeof(double_t) + (Nvissec*2+Nweightss)*sizeof(float_t);
double_t *memory = (double*) malloc ( (Nsec*3)*sizeof(double_t) +
(Nvissec*2+Nweightss)*sizeof(float_t) );
if ( memory == NULL )
shutdown_wstacking(NOT_ENOUGH_MEM_STACKING, "Not enough memory for stacking", __FILE__, __LINE__);
double_t *uus = (double*) memory;
double_t *vvs = (double*) uus+Nsec;
double_t *wws = (double*) vvs+Nsec;
float_t *weightss = (float_t*)(wws+Nsec);
float_t *visreals = weightss + Nweightss;
float_t *visimgs = visreals + Nvissec;
//Initialize nccl //Initialize nccl
...@@ -129,7 +105,23 @@ void gridding_data(){ ...@@ -129,7 +105,23 @@ void gridding_data(){
double start = CPU_TIME_wt; double start = CPU_TIME_wt;
memset( memory, 0, mem_size ); uint Nsec = histo_send[isector];
uint Nweightss = Nsec*metaData.polarisations;
uint Nvissec = Nweightss*metaData.freq_per_chan;
double_t *memory = (double*) malloc ( (Nsec*3)*sizeof(double_t) +
(Nvissec*2+Nweightss)*sizeof(float_t) );
if ( memory == NULL )
shutdown_wstacking(NOT_ENOUGH_MEM_STACKING, "Not enough memory for stacking", __FILE__, __LINE__);
double_t *uus = (double_t*) memory;
double_t *vvs = (double_t*) uus+Nsec;
double_t *wws = (double_t*) vvs+Nsec;
float_t *weightss = (float_t*)((double_t*)wws+Nsec);
float_t *visreals = (float_t*)weightss + Nweightss;
float_t *visimgs = (float_t*)visreals + Nvissec;
// select data for this sector // select data for this sector
uint icount = 0; uint icount = 0;
...@@ -246,10 +238,10 @@ void gridding_data(){ ...@@ -246,10 +238,10 @@ void gridding_data(){
// int target_rank = (int)isector; it implied that size >= nsectors // int target_rank = (int)isector; it implied that size >= nsectors
int target_rank = (int)(isector % size); int target_rank = (int)(isector % size);
start = CPU_TIME_wt;
cudaStreamSynchronize(stream_reduce); cudaStreamSynchronize(stream_reduce);
start = CPU_TIME_wt;
ncclReduce(gridss_gpu, grid_gpu, size_of_grid, ncclDouble, ncclSum, target_rank, comm, stream_reduce); ncclReduce(gridss_gpu, grid_gpu, size_of_grid, ncclDouble, ncclSum, target_rank, comm, stream_reduce);
cudaStreamSynchronize(stream_reduce); cudaStreamSynchronize(stream_reduce);
...@@ -260,11 +252,12 @@ void gridding_data(){ ...@@ -260,11 +252,12 @@ void gridding_data(){
memset ( gridss, 0, 2*param.num_w_planes*xaxis*yaxis * sizeof(double) ); memset ( gridss, 0, 2*param.num_w_planes*xaxis*yaxis * sizeof(double) );
} }
free(memory);
} }
//Copy data back from device to host (to be deleted in next steps) //Copy data back from device to host (to be deleted in next steps)
free(memory);
cudaMemcpyAsync(grid, grid_gpu, 2*param.num_w_planes*xaxis*yaxis*sizeof(double), cudaMemcpyDeviceToHost, stream_reduce); cudaMemcpyAsync(grid, grid_gpu, 2*param.num_w_planes*xaxis*yaxis*sizeof(double), cudaMemcpyDeviceToHost, stream_reduce);
MPI_Barrier(MPI_COMM_WORLD); MPI_Barrier(MPI_COMM_WORLD);
......
gridding_rccl.cpp
+ 22
− 31
View file @ ea94cf82
...@@ -64,32 +64,6 @@ void gridding_data(){ ...@@ -64,32 +64,6 @@ void gridding_data(){
if ( rank == 0 ) if ( rank == 0 )
printf("RESOLUTION = %f rad, %f arcsec\n", resolution, resolution_asec); printf("RESOLUTION = %f rad, %f arcsec\n", resolution, resolution_asec);
// find the largest value in histo_send[]
//
uint Nsec = histo_send[0];
for (uint isector = 1; isector < nsectors; isector++)
Nsec = ( Nsec < histo_send[isector] ? histo_send[isector] : Nsec );
uint Nweightss = Nsec*metaData.polarisations;
uint Nvissec = Nweightss*metaData.freq_per_chan;
// allocate sector arrays
// note: we use the largest allocation among all sectors
//
unsigned long long int mem_size = (Nsec*3)*sizeof(double_t) + (Nvissec*2+Nweightss)*sizeof(float_t);
double_t *memory = (double*) malloc ( (Nsec*3)*sizeof(double_t) +
(Nvissec*2+Nweightss)*sizeof(float_t) );
if ( memory == NULL )
shutdown_wstacking(NOT_ENOUGH_MEM_STACKING, "Not enough memory for stacking", __FILE__, __LINE__);
double_ty *uus = (double*) memory;
double_ty *vvs = (double*) uus+Nsec;
double_ty *wws = (double*) vvs+Nsec;
float_ty *weightss = (float_ty*)(wws+Nsec);
float_ty *visreals = weightss + Nweightss;
float_ty *visimgs = visreals + Nvissec;
//Initialize nccl //Initialize nccl
...@@ -127,7 +101,23 @@ void gridding_data(){ ...@@ -127,7 +101,23 @@ void gridding_data(){
double start = CPU_TIME_wt; double start = CPU_TIME_wt;
memset( memory, 0, mem_size ); uint Nsec = histo_send[isector];
uint Nweightss = Nsec*metaData.polarisations;
uint Nvissec = Nweightss*metaData.freq_per_chan;
double_t *memory = (double*) malloc ( (Nsec*3)*sizeof(double_t) +
(Nvissec*2+Nweightss)*sizeof(float_t) );
if ( memory == NULL )
shutdown_wstacking(NOT_ENOUGH_MEM_STACKING, "Not enough memory for stacking", __FILE__, __LINE__);
double_t *uus = (double_t*) memory;
double_t *vvs = (double_t*) uus+Nsec;
double_t *wws = (double_t*) vvs+Nsec;
float_t *weightss = (float_t*)((double_t*)wws+Nsec);
float_t *visreals = (float_t*)weightss + Nweightss;
float_t *visimgs = (float_t*)visreals + Nvissec;
// select data for this sector // select data for this sector
uint icount = 0; uint icount = 0;
...@@ -244,10 +234,11 @@ void gridding_data(){ ...@@ -244,10 +234,11 @@ void gridding_data(){
// int target_rank = (int)isector; it implied that size >= nsectors // int target_rank = (int)isector; it implied that size >= nsectors
int target_rank = (int)(isector % size); int target_rank = (int)(isector % size);
start = CPU_TIME_wt;
hipStreamSynchronize(stream_reduce); hipStreamSynchronize(stream_reduce);
start = CPU_TIME_wt;
ncclReduce(gridss_gpu, grid_gpu, size_of_grid, ncclDouble, ncclSum, target_rank, comm, stream_reduce); ncclReduce(gridss_gpu, grid_gpu, size_of_grid, ncclDouble, ncclSum, target_rank, comm, stream_reduce);
hipStreamSynchronize(stream_reduce); hipStreamSynchronize(stream_reduce);
...@@ -258,11 +249,11 @@ void gridding_data(){ ...@@ -258,11 +249,11 @@ void gridding_data(){
memset ( gridss, 0, 2*param.num_w_planes*xaxis*yaxis * sizeof(double) ); memset ( gridss, 0, 2*param.num_w_planes*xaxis*yaxis * sizeof(double) );
} }
free(memory);
} }
//Copy data back from device to host (to be deleted in next steps) //Copy data back from device to host (to be deleted in next steps)
free(memory);
hipMemcpyAsync(grid, grid_gpu, 2*param.num_w_planes*xaxis*yaxis*sizeof(double), hipMemcpyDeviceToHost, stream_reduce); hipMemcpyAsync(grid, grid_gpu, 2*param.num_w_planes*xaxis*yaxis*sizeof(double), hipMemcpyDeviceToHost, stream_reduce);
MPI_Barrier(MPI_COMM_WORLD); MPI_Barrier(MPI_COMM_WORLD);
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment