You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

735 lines
21 KiB

/*
* Test GPU aware MPI on different platforms using a distributed
* 1d stencil on a 2d array. The exchange in second (non-contiguous)
* direction forces use of staging buffers, which replicates what
* is needed for all but the innermost dimension exchanges in the
* GENE fusion code.
*
* Takes optional command line arg for size of each dimension of the domain
* n_global, in 1024 increments. Default is 8 * 1024 (so 256K plus ghost points
* in size for doulbles per array), which should fit on any system but may not
* be enough to tax larger HPC GPUs and MPI impelmentations.
*
* There will be four exchange buffers of size 2 * n_global, i.e. 128K each
* by default.
*
* Gtensor is used so a single source can be used for all platforms.
*/
#include <cmath>
#include <mpi.h>
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include "gtensor/gtensor.h"
#include "gtensor/reductions.h"
using namespace gt::placeholders;
const double PI = 3.141592653598793;
inline void check(const char* fname, int line, int mpi_rval)
{
if (mpi_rval != MPI_SUCCESS) {
printf("%s:%d error %d\n", fname, line, mpi_rval);
exit(2);
}
}
#define CHECK(x) check(__FILE__, __LINE__, (x))
// little hack to make code parameterizable on managed vs device memory
namespace gt
{
namespace ext
{
namespace detail
{
template <typename T, gt::size_type N, typename S = gt::space::device>
struct gthelper
{
using gtensor = gt::gtensor<T, N, S>;
};
#ifdef GTENSOR_HAVE_DEVICE
template <typename T, gt::size_type N>
struct gthelper<T, N, gt::space::managed>
{
using gtensor = gt::gtensor_container<gt::space::managed_vector<T>, N>;
};
#endif
} // namespace detail
template <typename T, gt::size_type N, typename S = gt::space::device>
using gtensor2 = typename detail::gthelper<T, N, S>::gtensor;
} // namespace ext
} // namespace gt
static const gt::gtensor<double, 1> stencil5 = {1.0 / 12.0, -2.0 / 3.0, 0.0,
2.0 / 3.0, -1.0 / 12.0};
/*
* Return unevaluated expression that calculates the 1d stencil in the
* first dimension of a 2d array.
*
* Size of the result will be size of z with minus 4 in second dimension.
*/
template <typename S>
inline auto stencil2d_1d_5_d0(const gt::ext::gtensor2<double, 2, S>& z,
const gt::gtensor<double, 1>& stencil)
{
return stencil(0) * z.view(_s(0, -4), _all) +
stencil(1) * z.view(_s(1, -3), _all) +
stencil(2) * z.view(_s(2, -2), _all) +
stencil(3) * z.view(_s(3, -1), _all) +
stencil(4) * z.view(_s(4, _), _all);
}
/*
* Return unevaluated expression that calculates the 1d stencil in the
* second dimension of a 2d array.
*
* Size of the result will be size of z with minus 4 in second dimension.
*/
template <typename S>
inline auto stencil2d_1d_5_d1(const gt::ext::gtensor2<double, 2, S>& z,
const gt::gtensor<double, 1>& stencil)
{
return stencil(0) * z.view(_all, _s(0, -4)) +
stencil(1) * z.view(_all, _s(1, -3)) +
stencil(2) * z.view(_all, _s(2, -2)) +
stencil(3) * z.view(_all, _s(3, -1)) +
stencil(4) * z.view(_all, _s(4, _));
}
void set_rank_device(int n_ranks, int rank)
{
int n_devices, device, ranks_per_device;
n_devices = gt::backend::clib::device_get_count();
if (n_ranks > n_devices) {
if (n_ranks % n_devices != 0) {
printf(
"ERROR: Number of ranks (%d) not a multiple of number of GPUs (%d)\n",
n_ranks, n_devices);
exit(EXIT_FAILURE);
}
ranks_per_device = n_ranks / n_devices;
device = rank / ranks_per_device;
} else {
ranks_per_device = 1;
device = rank;
}
gt::backend::clib::device_set(device);
}
// exchange in first dimension, staging into contiguous buffers on device
template <typename S>
void boundary_exchange_x(MPI_Comm comm, int world_size, int rank,
gt::ext::gtensor2<double, 2, S>& d_z, int n_bnd,
bool stage_host = false)
{
auto buf_shape = gt::shape(n_bnd, d_z.shape(1));
gt::gtensor_device<double, 2> sbuf_l(buf_shape);
gt::gtensor_device<double, 2> sbuf_r(buf_shape);
gt::gtensor_device<double, 2> rbuf_r(buf_shape);
gt::gtensor_device<double, 2> rbuf_l(buf_shape);
gt::shape_type<2> host_buf_shape;
if (stage_host) {
host_buf_shape = buf_shape;
} else {
host_buf_shape = {0, 0};
}
gt::gtensor<double, 2> h_sbuf_l(host_buf_shape);
gt::gtensor<double, 2> h_sbuf_r(host_buf_shape);
gt::gtensor<double, 2> h_rbuf_r(host_buf_shape);
gt::gtensor<double, 2> h_rbuf_l(host_buf_shape);
MPI_Request req_l[2];
MPI_Request req_r[2];
int rank_l = rank - 1;
int rank_r = rank + 1;
// start async copy of ghost points into send buffers
if (rank_l >= 0) {
sbuf_l = d_z.view(_s(n_bnd, 2 * n_bnd), _all);
if (stage_host) {
gt::copy(sbuf_l, h_sbuf_l);
}
}
if (rank_r <= world_size) {
sbuf_r = d_z.view(_s(-2 * n_bnd, -n_bnd), _all);
if (stage_host) {
gt::copy(sbuf_r, h_sbuf_r);
}
}
// initiate async recv
if (rank_l >= 0) {
double* rbuf_l_data = nullptr;
if (stage_host) {
rbuf_l_data = h_rbuf_l.data();
} else {
rbuf_l_data = rbuf_l.data().get();
}
CHECK(MPI_Irecv(rbuf_l_data, rbuf_l.size(), MPI_DOUBLE, rank_l, 123, comm,
&req_l[0]));
}
if (rank_r < world_size) {
double* rbuf_r_data = nullptr;
if (stage_host) {
rbuf_r_data = h_rbuf_r.data();
} else {
rbuf_r_data = rbuf_r.data().get();
}
CHECK(MPI_Irecv(rbuf_r_data, rbuf_r.size(), MPI_DOUBLE, rank_r, 456, comm,
&req_r[0]));
}
// wait for send buffer fill
gt::synchronize();
// initiate async sends
if (rank_l >= 0) {
double* sbuf_l_data = nullptr;
if (stage_host) {
sbuf_l_data = h_sbuf_l.data();
} else {
sbuf_l_data = sbuf_l.data().get();
}
CHECK(MPI_Isend(sbuf_l_data, sbuf_l.size(), MPI_DOUBLE, rank_l, 456, comm,
&req_l[1]));
}
if (rank_r < world_size) {
double* sbuf_r_data = nullptr;
if (stage_host) {
sbuf_r_data = h_sbuf_r.data();
} else {
sbuf_r_data = sbuf_r.data().get();
}
CHECK(MPI_Isend(sbuf_r_data, sbuf_r.size(), MPI_DOUBLE, rank_r, 123, comm,
&req_r[1]));
}
// wait for send/recv to complete, then copy data back into main data array
int mpi_rval;
if (rank_l >= 0) {
MPI_Status status[2];
mpi_rval = MPI_Waitall(2, req_l, status);
if (mpi_rval != MPI_SUCCESS) {
printf("send_l error: %d (%d %d)\n", mpi_rval, status[0].MPI_ERROR,
status[1].MPI_ERROR);
}
if (stage_host) {
gt::copy(h_rbuf_l, rbuf_l);
}
d_z.view(_s(0, n_bnd), _all) = rbuf_l;
}
if (rank_r < world_size) {
MPI_Status status[2];
mpi_rval = MPI_Waitall(2, req_r, status);
if (mpi_rval != MPI_SUCCESS) {
printf("send_r error: %d (%d %d)\n", mpi_rval, status[0].MPI_ERROR,
status[1].MPI_ERROR);
}
if (stage_host) {
gt::copy(h_rbuf_r, rbuf_r);
}
d_z.view(_s(-n_bnd, _), _all) = rbuf_r;
}
gt::synchronize();
}
// exchange in second dimension, optional staging into device buffer
template <typename S>
void boundary_exchange_y(MPI_Comm comm, int world_size, int rank,
gt::ext::gtensor2<double, 2, S>& d_z, int n_bnd,
bool stage_device)
{
gt::shape_type<2> buf_shape;
if (stage_device) {
buf_shape = gt::shape(d_z.shape(0), n_bnd);
} else {
buf_shape = {0, 0};
}
gt::gtensor_device<double, 2> sbuf_l(buf_shape);
gt::gtensor_device<double, 2> sbuf_r(buf_shape);
gt::gtensor_device<double, 2> rbuf_r(buf_shape);
gt::gtensor_device<double, 2> rbuf_l(buf_shape);
MPI_Request req_l[2];
MPI_Request req_r[2];
int rank_l = rank - 1;
int rank_r = rank + 1;
auto sv_l = gt::view_strided(d_z, _all, _s(n_bnd, 2 * n_bnd));
auto sv_r = gt::view_strided(d_z, _all, _s(-2 * n_bnd, -n_bnd));
auto rv_l = gt::view_strided(d_z, _all, _s(0, n_bnd));
auto rv_r = gt::view_strided(d_z, _all, _s(-n_bnd, _));
// start async copy of ghost points into send buffers
if (stage_device) {
if (rank_l >= 0) {
sbuf_l = sv_l;
}
if (rank_r <= world_size) {
sbuf_r = sv_r;
}
}
// initiate async recv
if (rank_l >= 0) {
double* rbuf_l_data = nullptr;
if (stage_device) {
rbuf_l_data = rbuf_l.data().get();
} else {
rbuf_l_data = rv_l.data().get();
}
CHECK(MPI_Irecv(rbuf_l_data, rv_l.size(), MPI_DOUBLE, rank_l, 123, comm,
&req_l[0]));
}
if (rank_r < world_size) {
double* rbuf_r_data = nullptr;
if (stage_device) {
rbuf_r_data = rbuf_r.data().get();
} else {
rbuf_r_data = rv_r.data().get();
}
CHECK(MPI_Irecv(rbuf_r_data, rv_r.size(), MPI_DOUBLE, rank_r, 456, comm,
&req_r[0]));
}
// wait for send buffer fill
// if (stage_device) {
gt::synchronize();
// }
// initiate async sends
if (rank_l >= 0) {
double* sbuf_l_data = nullptr;
if (stage_device) {
sbuf_l_data = sbuf_l.data().get();
} else {
sbuf_l_data = sv_l.data().get();
}
CHECK(MPI_Isend(sbuf_l_data, sv_l.size(), MPI_DOUBLE, rank_l, 456, comm,
&req_l[1]));
}
if (rank_r < world_size) {
double* sbuf_r_data = nullptr;
if (stage_device) {
sbuf_r_data = sbuf_r.data().get();
} else {
sbuf_r_data = sv_r.data().get();
}
CHECK(MPI_Isend(sbuf_r_data, sv_r.size(), MPI_DOUBLE, rank_r, 123, comm,
&req_r[1]));
}
// wait for send/recv to complete, then copy data back into main data array
int mpi_rval;
if (rank_l >= 0) {
MPI_Status status[2];
mpi_rval = MPI_Waitall(2, req_l, status);
if (mpi_rval != MPI_SUCCESS) {
printf("send_l error: %d (%d %d)\n", mpi_rval, status[0].MPI_ERROR,
status[1].MPI_ERROR);
}
if (stage_device) {
gt::copy(rbuf_l, rv_l);
}
}
if (rank_r < world_size) {
MPI_Status status[2];
mpi_rval = MPI_Waitall(2, req_r, status);
if (mpi_rval != MPI_SUCCESS) {
printf("send_r error: %d (%d %d)\n", mpi_rval, status[0].MPI_ERROR,
status[1].MPI_ERROR);
}
if (stage_device) {
gt::copy(rbuf_r, rv_r);
}
}
gt::synchronize();
}
template <int Dim, typename S>
void print_test_name(bool use_buffers)
{
if constexpr (std::is_same<S, gt::space::device>::value) {
printf("TEST dim:%d, device , buf:%d", Dim, use_buffers);
} else {
printf("TEST dim:%d, managed, buf:%d", Dim, use_buffers);
}
}
template <typename S, int Dim>
void test_deriv(int device_id, uint32_t vendor_id, int world_size,
int world_rank, int n_global, int n_iter, bool use_buffers,
int n_warmup = 5)
{
// Note: domain will be n_global x n_global plus ghost points in one dimension
int n_sten = 5;
int n_bnd = (n_sten - 1) / 2;
const int n_local = n_global / world_size;
int nx_local, ny_local;
int nx_local_ghost, ny_local_ghost;
int nx_bnd, ny_bnd;
if constexpr (Dim == 0) {
nx_bnd = n_bnd;
ny_bnd = 0;
nx_local = n_local;
nx_local_ghost = n_local + 2 * n_bnd;
ny_local = n_global;
ny_local_ghost = n_global;
} else {
nx_bnd = 0;
ny_bnd = n_bnd;
nx_local = n_global;
nx_local_ghost = n_global;
ny_local = n_local;
ny_local_ghost = n_local + 2 * n_bnd;
}
gt::shape_type<2> z_shape(nx_local_ghost, ny_local_ghost);
gt::shape_type<2> dz_shape(nx_local, ny_local);
auto h_z = gt::empty<double>(z_shape);
gt::ext::gtensor2<double, 2, S> d_z(z_shape);
auto h_dz_numeric = gt::empty<double>(dz_shape);
auto h_dz_actual = gt::empty<double>(dz_shape);
gt::ext::gtensor2<double, 2, S> d_dz_numeric(dz_shape);
double ln = 8;
double delta = ln / n_global;
double ln_local = ln / world_size;
double scale = n_global / ln;
auto fn = [](double x, double y) { return x * x * x + y * y; };
auto fn_dzdx = [](double x, double y) { return 3 * x * x; };
auto fn_dzdy = [](double x, double y) { return 2 * y; };
struct timespec start, end;
double iter_time = 0.0;
double total_time = 0.0;
double x_start = 0, y_start = 0;
if constexpr (Dim == 0) {
x_start = world_rank * ln_local;
} else {
y_start = world_rank * ln_local;
}
for (int j = 0; j < ny_local; j++) {
double ytmp = y_start + j * delta;
for (int i = 0; i < nx_local; i++) {
double xtmp = x_start + i * delta;
h_z(i + nx_bnd, j + ny_bnd) = fn(xtmp, ytmp);
if constexpr (Dim == 0) {
h_dz_actual(i, j) = fn_dzdx(xtmp, ytmp);
} else {
h_dz_actual(i, j) = fn_dzdy(xtmp, ytmp);
}
}
}
// fill boundary points on ends
if constexpr (Dim == 0) {
if (world_rank == 0) {
for (int j = 0; j < ny_local; j++) {
double ytmp = j * delta;
for (int i = 0; i < nx_bnd; i++) {
double xtmp = (i - nx_bnd) * delta;
h_z(i, j) = fn(xtmp, ytmp);
}
}
}
if (world_rank == world_size - 1) {
for (int j = 0; j < ny_local; j++) {
double ytmp = j * delta;
for (int i = 0; i < nx_bnd; i++) {
double xtmp = ln + i * delta;
h_z(nx_bnd + nx_local + i, j) = fn(xtmp, ytmp);
}
}
}
} else {
if (world_rank == 0) {
for (int j = 0; j < ny_bnd; j++) {
double ytmp = (j - ny_bnd) * delta;
for (int i = 0; i < nx_local; i++) {
double xtmp = i * delta;
h_z(i, j) = fn(xtmp, ytmp);
}
}
}
if (world_rank == world_size - 1) {
for (int j = 0; j < ny_bnd; j++) {
double ytmp = ln + j * delta;
for (int i = 0; i < nx_local; i++) {
double xtmp = i * delta;
h_z(i, ny_bnd + ny_local + j) = fn(xtmp, ytmp);
}
}
}
}
/*
for (int i = 0; i < 5; i++) {
printf("%d row1-l %f\n", world_rank, h_z(1, i));
}
for (int i = 0; i < 5; i++) {
printf("%d row1-r %f\n", world_rank, h_z(1, n_local_with_ghost - 1 - i));
}
*/
gt::copy(h_z, d_z);
// gt::synchronize();
for (int i = 0; i < n_warmup + n_iter; i++) {
clock_gettime(CLOCK_MONOTONIC, &start);
if constexpr (Dim == 0) {
boundary_exchange_x<S>(MPI_COMM_WORLD, world_size, world_rank, d_z, n_bnd,
use_buffers);
} else {
boundary_exchange_y<S>(MPI_COMM_WORLD, world_size, world_rank, d_z, n_bnd,
use_buffers);
}
clock_gettime(CLOCK_MONOTONIC, &end);
iter_time =
((end.tv_sec - start.tv_sec) + (end.tv_nsec - start.tv_nsec) * 1.0e-9);
if (i >= n_warmup) {
total_time += iter_time;
}
// do some calculation, to try to more closely simulate what happens in GENE
if constexpr (Dim == 0) {
d_dz_numeric = stencil2d_1d_5_d0<S>(d_z, stencil5) * scale;
} else {
d_dz_numeric = stencil2d_1d_5_d1<S>(d_z, stencil5) * scale;
}
gt::synchronize();
}
#ifdef DEBUG
printf("%d/%d exchange time %0.8f ms\n", world_rank, world_size,
total_time / n_iter * 1000);
#endif
gt::copy(d_dz_numeric, h_dz_numeric);
/*
for (int i = 0; i < 5; i++) {
printf("%d la %f\n%d ln %f\n", world_rank, h_dzdx_actual(8, i), world_rank,
h_dzdx_numeric(8, i));
}
for (int i = 0; i < 5; i++) {
int idx = n_local - 1 - i;
printf("%d ra %f\n%d rn %f\n", world_rank, h_dzdx_actual(8, idx),
world_rank, h_dzdx_numeric(8, idx));
}
*/
double err_norm = std::sqrt(gt::sum_squares(h_dz_numeric - h_dz_actual));
#ifdef DEBUG
printf("%d/%d [%d:0x%08x] err_norm = %.8f\n", world_rank, world_size,
device_id, vendor_id, err_norm);
#endif
double time_sum;
MPI_Reduce(&total_time, &time_sum, 1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
double err_sum;
MPI_Reduce(&err_norm, &err_sum, 1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
if (world_rank == 0) {
print_test_name<Dim, S>(use_buffers);
printf("; %0.8f, err=%0.8f\n", time_sum, err_sum);
}
}
template <typename S, int Dim>
void test_sum(int device_id, uint32_t vendor_id, int world_size, int world_rank,
int n_global, int n_iter, int n_warmup = 5)
{
// Note: domain will be n_global x n_global plus ghost points in one dimension
const int n_local = n_global / world_size;
int nx_local, ny_local;
struct timespec start, end;
double iter_time = 0.0;
double total_time = 0.0;
if constexpr (Dim == 0) {
nx_local = n_local;
ny_local = n_global;
} else {
nx_local = n_global;
ny_local = n_local;
}
gt::shape_type<2> z_shape(nx_local, ny_local);
gt::ext::gtensor2<double, 2, S> d_z(z_shape, PI / world_size);
// reduction test
gt::shape_type<1> sum_shape;
if constexpr (Dim == 0) {
sum_shape = gt::shape(d_z.shape(0));
} else {
sum_shape = gt::shape(d_z.shape(1));
}
gt::ext::gtensor2<double, 1, S> d_sum(sum_shape);
gt::gtensor<double, 1> h_sum(sum_shape);
for (int i = 0; i < n_warmup + n_iter; i++) {
if constexpr (Dim == 0) {
gt::sum_axis_to(d_sum, d_z, 0);
gt::synchronize();
clock_gettime(CLOCK_MONOTONIC, &start);
CHECK(MPI_Allreduce(MPI_IN_PLACE, d_sum.data().get(), d_sum.size(),
MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD));
clock_gettime(CLOCK_MONOTONIC, &end);
} else {
gt::sum_axis_to(d_sum, d_z, 1);
gt::synchronize();
clock_gettime(CLOCK_MONOTONIC, &start);
CHECK(MPI_Allreduce(MPI_IN_PLACE, d_sum.data().get(), d_sum.size(),
MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD));
clock_gettime(CLOCK_MONOTONIC, &end);
}
iter_time =
((end.tv_sec - start.tv_sec) + (end.tv_nsec - start.tv_nsec) * 1.0e-9);
if (i >= n_warmup) {
total_time += iter_time;
}
}
#ifdef DEBUG
printf("%d/%d allreduce time %0.8f ms\n", world_rank, world_size,
total_time / n_iter * 1000);
#endif
gt::copy(d_sum, h_sum);
double time_sum;
MPI_Reduce(&total_time, &time_sum, 1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
if (world_rank == 0) {
print_test_name<Dim, S>(false);
printf("; allreduce=%0.8f\n", time_sum);
}
}
int main(int argc, char** argv)
{
using S = gt::space::managed;
// Note: domain will be n_global x n_global plus ghost points in one dimension
int n_global = 8 * 1024;
int n_iter = 1000;
int n_warmup = 10;
if (argc > 1) {
n_global = std::atoi(argv[1]) * 1024;
}
if (argc > 2) {
n_iter = std::atoi(argv[2]);
}
int world_size, world_rank, device_id;
uint32_t vendor_id;
CHECK(MPI_Init(NULL, NULL));
CHECK(MPI_Comm_size(MPI_COMM_WORLD, &world_size));
CHECK(MPI_Comm_rank(MPI_COMM_WORLD, &world_rank));
if (n_global % world_size != 0) {
printf("%d nmpi (%d) must be divisor of domain size (%d), exiting\n",
world_rank, world_size, n_global);
exit(1);
}
const int n_local = n_global / world_size;
set_rank_device(world_size, world_rank);
device_id = gt::backend::clib::device_get();
vendor_id = gt::backend::clib::device_get_vendor_id(device_id);
if (world_rank == 0) {
printf("n procs = %d\n", world_size);
printf("n_global = %d\n", n_global);
printf("n_local = %d\n", n_local);
printf("n_iter = %d\n", n_iter);
printf("n_warmup = %d\n", n_warmup);
}
fflush(stdout);
test_deriv<gt::space::device, 0>(device_id, vendor_id, world_size, world_rank,
n_global, n_iter, true, 5);
test_deriv<gt::space::device, 0>(device_id, vendor_id, world_size, world_rank,
n_global, n_iter, false, 5);
#ifdef TEST_MANAGED
test_deriv<gt::space::managed, 0>(device_id, vendor_id, world_size,
world_rank, n_global, n_iter, true, 5);
test_deriv<gt::space::managed, 0>(device_id, vendor_id, world_size,
world_rank, n_global, n_iter, false, 5);
#endif
test_deriv<gt::space::device, 1>(device_id, vendor_id, world_size, world_rank,
n_global, n_iter, true, 5);
test_deriv<gt::space::device, 1>(device_id, vendor_id, world_size, world_rank,
n_global, n_iter, false, 5);
#ifdef TEST_MANAGED
test_deriv<gt::space::managed, 1>(device_id, vendor_id, world_size,
world_rank, n_global, n_iter, true, 5);
test_deriv<gt::space::managed, 1>(device_id, vendor_id, world_size,
world_rank, n_global, n_iter, false, 5);
#endif
test_sum<gt::space::device, 0>(device_id, vendor_id, world_size, world_rank,
n_global, n_iter, 5);
#ifdef TEST_MANAGED
test_sum<gt::space::managed, 0>(device_id, vendor_id, world_size, world_rank,
n_global, n_iter, 5);
#endif
test_sum<gt::space::device, 1>(device_id, vendor_id, world_size, world_rank,
n_global, n_iter, 5);
#ifdef TEST_MANAGED
test_sum<gt::space::managed, 1>(device_id, vendor_id, world_size, world_rank,
n_global, n_iter, 5);
#endif
MPI_Finalize();
return EXIT_SUCCESS;
}