Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Remove managed memory from read_workerflow #82

Closed
wants to merge 2 commits into from
Closed

Remove managed memory from read_workerflow #82

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e8b5b2d
Remove managed memory from read_workerflow
atmyers Sep 13, 2024
011d4d4
run assignTeachersAndWorkgroup on the host for now
atmyers Sep 13, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 0 additions & 1 deletion src/CensusData.H
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -50,7 +50,6 @@ struct CensusData
std::vector<int> num_cases,
const std::vector<std::string>& names);

private:
void assignTeachersAndWorkgroup (AgentContainer& pc);

};
Expand Down
93 changes: 62 additions & 31 deletions src/CensusData.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -546,10 +546,10 @@ void CensusData::read_workerflow (AgentContainer& pc, /*!< Agent conta
{
/* Allocate worker-flow matrix, only from units with nighttime
communities on this processor (Unit_on_proc[] flag) */
unsigned int** flow = (unsigned int **) amrex::The_Arena()->alloc(demo.Nunit*sizeof(unsigned int *));
unsigned int** flow = (unsigned int **) amrex::The_Pinned_Arena()->alloc(demo.Nunit*sizeof(unsigned int *));
for (int i = 0; i < demo.Nunit; i++) {
if (demo.Unit_on_proc[i]) {
flow[i] = (unsigned int *) amrex::The_Arena()->alloc(demo.Nunit*sizeof(unsigned int));
flow[i] = (unsigned int *) amrex::The_Pinned_Arena()->alloc(demo.Nunit*sizeof(unsigned int));
for (int j = 0; j < demo.Nunit; j++) flow[i][j] = 0;
}
}
Expand Down Expand Up @@ -611,6 +611,28 @@ void CensusData::read_workerflow (AgentContainer& pc, /*!< Agent conta
}
}

unsigned int** d_flow = (unsigned int **) amrex::The_Device_Arena()->alloc(demo.Nunit*sizeof(unsigned int *));
amrex::Gpu::HostVector<unsigned int*> host_vector_flow(demo.Nunit, nullptr);
for (int i = 0; i < demo.Nunit; i++) {
if (demo.Unit_on_proc[i]) {
host_vector_flow[i] = (unsigned int *) amrex::The_Device_Arena()->alloc(demo.Nunit*sizeof(unsigned int));
}
}

amrex::Gpu::copy(Gpu::hostToDevice, host_vector_flow.begin(), host_vector_flow.end(), d_flow);
for (int i = 0; i < demo.Nunit; i++) {
if (demo.Unit_on_proc[i]) {
amrex::Gpu::copy(Gpu::hostToDevice, flow[i], flow[i] + demo.Nunit, host_vector_flow[i]);
}
}

for (int i = 0; i < demo.Nunit; i++) {
if (demo.Unit_on_proc[i]) {
amrex::The_Pinned_Arena()->free(flow[i]);
}
}
amrex::The_Pinned_Arena()->free(flow);

const Box& domain = pc.Geom(0).Domain();

/* This is where workplaces should be assigned */
Expand Down Expand Up @@ -655,10 +677,10 @@ void CensusData::read_workerflow (AgentContainer& pc, /*!< Agent conta
unsigned int irnd = amrex::Random_int(nwork, engine);
int to = 0;
int comm_to = 0;
if (irnd < flow[from][Nunit-1]) {
if (irnd < d_flow[from][Nunit-1]) {
/* Choose a random destination unit */
to = 0;
while (irnd >= flow[from][to]) { to++; }
while (irnd >= d_flow[from][to]) { to++; }
}

/*If from=to unit, 25% EXTRA chance of working in home community*/
Expand All @@ -684,6 +706,14 @@ void CensusData::read_workerflow (AgentContainer& pc, /*!< Agent conta
}
});
}

for (int i = 0; i < demo.Nunit; i++) {
if (demo.Unit_on_proc[i]) {
amrex::The_Device_Arena()->free(host_vector_flow[i]);
}
}
amrex::The_Device_Arena()->free(d_flow);

assignTeachersAndWorkgroup(pc);
}

Expand All @@ -692,30 +722,30 @@ void CensusData::assignTeachersAndWorkgroup (AgentContainer& pc /*!< Agent
{
const Box& domain = pc.Geom(0).Domain();

auto total_teacher_unit = pc.getUnitTeacherCounts();
auto total_teacher_unit = pc.getUnitTeacherCounts().data();

auto total_teacher_counts = pc.getCommTeacherCounts();
amrex::Gpu::DeviceVector<int> total_teacher_counts_mod(total_teacher_counts.size(),0);
auto total_teacher_counts_ptr = total_teacher_counts.data();

auto daycr_teacher_counts = pc.getCommDayCrTeacherCounts();
amrex::Gpu::DeviceVector<int> daycr_teacher_counts_mod(daycr_teacher_counts.size(),0);
auto daycr_teacher_counts = pc.getCommDayCrTeacherCounts().data();
amrex::Gpu::DeviceVector<int> daycr_teacher_counts_mod(pc.getCommDayCrTeacherCounts().size(),0);
auto daycr_teacher_counts_ptr = daycr_teacher_counts_mod.data();

auto high_teacher_counts = pc.getCommHighTeacherCounts();
amrex::Gpu::DeviceVector<int> high_teacher_counts_mod(high_teacher_counts.size(),0);
auto high_teacher_counts = pc.getCommHighTeacherCounts().data();
amrex::Gpu::DeviceVector<int> high_teacher_counts_mod(pc.getCommHighTeacherCounts().size(),0);
auto high_teacher_counts_ptr = high_teacher_counts_mod.data();

auto middle_teacher_counts = pc.getCommMiddleTeacherCounts();
amrex::Gpu::DeviceVector<int> middle_teacher_counts_mod(middle_teacher_counts.size(),0);
auto middle_teacher_counts = pc.getCommMiddleTeacherCounts().data();
amrex::Gpu::DeviceVector<int> middle_teacher_counts_mod(pc.getCommMiddleTeacherCounts().size(),0);
auto middle_teacher_counts_ptr = middle_teacher_counts_mod.data();

auto elem3_teacher_counts = pc.getCommElem3TeacherCounts();
amrex::Gpu::DeviceVector<int> elem3_teacher_counts_mod(elem3_teacher_counts.size(),0);
auto elem3_teacher_counts = pc.getCommElem3TeacherCounts().data();
amrex::Gpu::DeviceVector<int> elem3_teacher_counts_mod(pc.getCommElem3TeacherCounts().size(),0);
auto elem3_teacher_counts_ptr = elem3_teacher_counts_mod.data();

auto elem4_teacher_counts = pc.getCommElem4TeacherCounts();
amrex::Gpu::DeviceVector<int> elem4_teacher_counts_mod(elem4_teacher_counts.size(),0);
auto elem4_teacher_counts = pc.getCommElem4TeacherCounts().data();
amrex::Gpu::DeviceVector<int> elem4_teacher_counts_mod(pc.getCommElem4TeacherCounts().size(),0);
auto elem4_teacher_counts_ptr = elem4_teacher_counts_mod.data();

#ifdef AMREX_USE_OMP
Expand All @@ -736,16 +766,17 @@ void CensusData::assignTeachersAndWorkgroup (AgentContainer& pc /*!< Agent
auto Ncommunity = demo.Ncommunity;

auto np = soa.numParticles();
for (int ip = 0; ip < np; ++ip) {

amrex::ParallelForRNG( np,
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This change is what causes the difference between the GPU and CPU results. Maybe there are updates in the loop that require atomics for the GPU

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, you're right - I think that instead of running this the Gpu, for now we need to explicitly copy the data off the device, run this loop in serial, then copy the data back. Since this only happens at initialization I think this will be okay.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

See update.

Copy link
Member Author

@atmyers atmyers Sep 13, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Note - with this change, AND with constructing the bins on the GPU, the code runs without crashing on development with amrex.the_arena_is_managed=0.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If I run unmanaged with random initial cases, I still get a segfault here, on line 961 in CensusData.cpp:
auto inds = bins.permutationPtr();

Copy link
Collaborator

@stevenhofmeyr stevenhofmeyr Sep 13, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This appears to go away if we use the GPU binning policy rather than Serial.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Binning with BinPolicy::Serial still relies on managed memory, but if you change that to BinPolicy::GPU, it works for me - is that not what you're seeing?

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, that's what i'm seeing. With GPU policy, it works. So we should keep the default managed memory for reproducibility. But I can now check the other branch to see if there is anywhere managed memory is being implicitly used.

[=] AMREX_GPU_DEVICE (int ip, RandomEngine const& engine) noexcept
{
int comm_to = (int) domain.index(IntVect(AMREX_D_DECL(work_i_ptr[ip],work_j_ptr[ip],0)));
if (comm_to >= Ncommunity) {
continue;
return;
}
int to = 0;
while (comm_to >= Start[to+1]) { to++; }

if (total_teacher_unit.data()[to] && (age_group_ptr[ip] == 2 || age_group_ptr[ip] == 3) && workgroup_ptr[ip] > 0)
if (total_teacher_unit[to] && (age_group_ptr[ip] == 2 || age_group_ptr[ip] == 3) && workgroup_ptr[ip] > 0)
{
int elem3_teacher = elem3_teacher_counts_ptr[comm_to];
int elem4_teacher = elem4_teacher_counts_ptr[comm_to];
Expand All @@ -755,22 +786,22 @@ void CensusData::assignTeachersAndWorkgroup (AgentContainer& pc /*!< Agent
int total = total_teacher_counts_ptr[comm_to];

// 50% chance of being a teacher if in working-age population (until max_teacher_numb is met)
if (amrex::Random() < 0.50 && (elem3_teacher + elem4_teacher
+ middle_teacher + high_teacher
+ daycr_teacher) < total)
if (amrex::Random(engine) < 0.50 && (elem3_teacher + elem4_teacher
+ middle_teacher + high_teacher
+ daycr_teacher) < total)
{
int available_slots[5] = {
elem3_teacher < elem3_teacher_counts.data()[comm_to],
elem4_teacher < elem4_teacher_counts.data()[comm_to],
middle_teacher < middle_teacher_counts.data()[comm_to],
high_teacher < high_teacher_counts.data()[comm_to],
daycr_teacher < daycr_teacher_counts.data()[comm_to]
elem3_teacher < elem3_teacher_counts[comm_to],
elem4_teacher < elem4_teacher_counts[comm_to],
middle_teacher < middle_teacher_counts[comm_to],
high_teacher < high_teacher_counts[comm_to],
daycr_teacher < daycr_teacher_counts[comm_to]
};

int total_available = available_slots[0] + available_slots[1] + available_slots[2] + available_slots[3] + available_slots[4];
if (total_available > 0)
{
int choice = amrex::Random_int(total_available);
int choice = amrex::Random_int(total_available, engine);
if (choice < available_slots[0]) {
school_ptr[ip] = 3; // elementary school for kids in Neighbordhood 1 & 2
workgroup_ptr[ip] = 3 ;
Expand Down Expand Up @@ -801,18 +832,18 @@ void CensusData::assignTeachersAndWorkgroup (AgentContainer& pc /*!< Agent
}
else{
constexpr int WG_size = 20;
unsigned int number = (unsigned int) rint( ((Real) Ndaywork[to] - total_teacher_unit.data()[to] ) /
unsigned int number = (unsigned int) rint( ((Real) Ndaywork[to] - total_teacher_unit[to] ) /
((Real) WG_size * (Start[to+1] - Start[to])) );

if (number) {
workgroup_ptr[ip] = 6 + amrex::Random_int(number);
workgroup_ptr[ip] = 6 + amrex::Random_int(number, engine);
work_nborhood_ptr[ip] = workgroup_ptr[ip] % 4; // each workgroup is assigned to a neighborhood as well
}

}
}

}
});
}
}

Expand Down
Loading