From c9f3c097ed5bce9adcc3f257b55a500e4dd9b90f Mon Sep 17 00:00:00 2001
From: Rafael Ravedutti <rafaelravedutti@gmail.com>
Date: Fri, 1 Dec 2023 19:20:00 +0100
Subject: [PATCH] Pack, communicate and unpack everything together in
synchronize()
Signed-off-by: Rafael Ravedutti <rafaelravedutti@gmail.com>
---
runtime/domain/regular_6d_stencil.cpp | 46 +++++++--
runtime/domain/regular_6d_stencil.hpp | 5 +
runtime/pairs.cpp | 74 ++++++--------
runtime/pairs.hpp | 6 ++
src/pairs/sim/comm.py | 136 +++++++++++++++++++++++---
5 files changed, 201 insertions(+), 66 deletions(-)
diff --git a/runtime/domain/regular_6d_stencil.cpp b/runtime/domain/regular_6d_stencil.cpp
index 4e7631f..f8f2472 100644
--- a/runtime/domain/regular_6d_stencil.cpp
+++ b/runtime/domain/regular_6d_stencil.cpp
@@ -118,7 +118,6 @@ void Regular6DStencil::communicateData(
int dim, int elem_size,
const real_t *send_buf, const int *send_offsets, const int *nsend,
real_t *recv_buf, const int *recv_offsets, const int *nrecv) {
- //std::cout << "communicateData" << std::endl;
const real_t *send_prev = &send_buf[send_offsets[dim * 2 + 0] * elem_size];
const real_t *send_next = &send_buf[send_offsets[dim * 2 + 1] * elem_size];
@@ -126,10 +125,6 @@ void Regular6DStencil::communicateData(
real_t *recv_next = &recv_buf[recv_offsets[dim * 2 + 1] * elem_size];
if(prev[dim] != rank) {
- //std::cout << rank << ": send " << nsend[dim * 2 + 0] << " elems to " << prev[dim] << ", recv " << nrecv[dim * 2 + 0] << " elems from " << next[dim] << std::endl;
- //MPI_Send(send_prev, nsend[dim * 2 + 0] * elem_size, MPI_DOUBLE, prev[dim], 0, MPI_COMM_WORLD);
- //MPI_Recv(recv_prev, nrecv[dim * 2 + 0] * elem_size, MPI_DOUBLE, next[dim], 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
-
MPI_Sendrecv(
send_prev, nsend[dim * 2 + 0] * elem_size, MPI_DOUBLE, prev[dim], 0,
recv_prev, nrecv[dim * 2 + 0] * elem_size, MPI_DOUBLE, next[dim], 0,
@@ -142,10 +137,6 @@ void Regular6DStencil::communicateData(
}
if(next[dim] != rank) {
- //std::cout << rank << ": send " << nsend[dim * 2 + 1] << " elems to " << next[dim] << ", recv " << nrecv[dim * 2 + 1] << " elems from " << prev[dim] << std::endl;
- //MPI_Send(send_next, nsend[dim * 2 + 1] * elem_size, MPI_DOUBLE, next[dim], 0, MPI_COMM_WORLD);
- //MPI_Recv(recv_next, nrecv[dim * 2 + 1] * elem_size, MPI_DOUBLE, prev[dim], 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
-
MPI_Sendrecv(
send_next, nsend[dim * 2 + 1] * elem_size, MPI_DOUBLE, next[dim], 0,
recv_next, nrecv[dim * 2 + 1] * elem_size, MPI_DOUBLE, prev[dim], 0,
@@ -158,4 +149,41 @@ void Regular6DStencil::communicateData(
}
}
+void Regular6DStencil::communicateAllData(
+ int ndims, int elem_size,
+ const real_t *send_buf, const int *send_offsets, const int *nsend,
+ real_t *recv_buf, const int *recv_offsets, const int *nrecv) {
+
+ for(int d = 0; d < ndims; d++) {
+ const real_t *send_prev = &send_buf[send_offsets[d * 2 + 0] * elem_size];
+ const real_t *send_next = &send_buf[send_offsets[d * 2 + 1] * elem_size];
+ real_t *recv_prev = &recv_buf[recv_offsets[d * 2 + 0] * elem_size];
+ real_t *recv_next = &recv_buf[recv_offsets[d * 2 + 1] * elem_size];
+
+ if(prev[d] != rank) {
+ MPI_Sendrecv(
+ send_prev, nsend[d * 2 + 0] * elem_size, MPI_DOUBLE, prev[d], 0,
+ recv_prev, nrecv[d * 2 + 0] * elem_size, MPI_DOUBLE, next[d], 0,
+ MPI_COMM_WORLD, MPI_STATUS_IGNORE);
+
+ } else {
+ for(int i = 0; i < nsend[d * 2 + 0] * elem_size; i++) {
+ recv_prev[i] = send_prev[i];
+ }
+ }
+
+ if(next[d] != rank) {
+ MPI_Sendrecv(
+ send_next, nsend[d * 2 + 1] * elem_size, MPI_DOUBLE, next[d], 0,
+ recv_next, nrecv[d * 2 + 1] * elem_size, MPI_DOUBLE, prev[d], 0,
+ MPI_COMM_WORLD, MPI_STATUS_IGNORE);
+
+ } else {
+ for(int i = 0; i < nsend[d * 2 + 1] * elem_size; i++) {
+ recv_next[i] = send_next[i];
+ }
+ }
+ }
+}
+
}
diff --git a/runtime/domain/regular_6d_stencil.hpp b/runtime/domain/regular_6d_stencil.hpp
index 1c12ff4..8a43b2d 100644
--- a/runtime/domain/regular_6d_stencil.hpp
+++ b/runtime/domain/regular_6d_stencil.hpp
@@ -54,6 +54,11 @@ public:
int dim, int elem_size,
const real_t *send_buf, const int *send_offsets, const int *nsend,
real_t *recv_buf, const int *recv_offsets, const int *nrecv);
+
+ void communicateAllData(
+ int ndims, int elem_size,
+ const real_t *send_buf, const int *send_offsets, const int *nsend,
+ real_t *recv_buf, const int *recv_offsets, const int *nrecv);
};
}
diff --git a/runtime/pairs.cpp b/runtime/pairs.cpp
index 9ec1f23..93c714c 100644
--- a/runtime/pairs.cpp
+++ b/runtime/pairs.cpp
@@ -305,7 +305,6 @@ void PairsSimulation::communicateSizes(int dim, const int *send_sizes, int *recv
array_flags->setHostFlag(nrecv_id);
array_flags->clearDeviceFlag(nrecv_id);
this->getDomainPartitioner()->communicateSizes(dim, send_sizes, recv_sizes);
- //PAIRS_DEBUG("send_sizes=[%d, %d], recv_sizes=[%d, %d]\n", send_sizes[dim * 2 + 0], send_sizes[dim * 2 + 1], recv_sizes[dim * 2 + 0], recv_sizes[dim * 2 + 1]);
}
void PairsSimulation::communicateData(
@@ -342,59 +341,42 @@ void PairsSimulation::communicateData(
dim, elem_size, send_buf, send_offsets, nsend, recv_buf, recv_offsets, nrecv);
copyArrayToDevice(recv_buf_id, Ignore, nrecv_all * elem_size * sizeof(real_t));
+}
- /*
- // Debug messages
- const int elems_to_print = 5;
-
- // Send buffer debug
- for(int i = 0; i < 2; i++) {
- int nsnd = nsend[dim * 2 + i];
-
- PAIRS_DEBUG("send_buf=[");
- for(int j = 0; j < MIN(elems_to_print, nsnd); j++) {
- for(int k = 0; k < elem_size; k++) {
- PAIRS_DEBUG("%f,", send_buf[(send_offsets[dim * 2 + i] + j) * elem_size + k]);
- }
- }
+void PairsSimulation::communicateAllData(
+ int ndims, int elem_size,
+ const real_t *send_buf, const int *send_offsets, const int *nsend,
+ real_t *recv_buf, const int *recv_offsets, const int *nrecv) {
- if(elems_to_print * 2 < nsnd) {
- PAIRS_DEBUG("\b ... ");
- }
+ auto send_buf_id = getArrayByHostPointer(send_buf).getId();
+ auto recv_buf_id = getArrayByHostPointer(recv_buf).getId();
+ auto send_offsets_id = getArrayByHostPointer(send_offsets).getId();
+ auto recv_offsets_id = getArrayByHostPointer(recv_offsets).getId();
+ auto nsend_id = getArrayByHostPointer(nsend).getId();
+ auto nrecv_id = getArrayByHostPointer(nrecv).getId();
- for(int j = MAX(elems_to_print, nsnd - elems_to_print); j < nsnd; j++) {
- for(int k = 0; k < elem_size; k++) {
- PAIRS_DEBUG("%f,", send_buf[(send_offsets[dim * 2 + i] + j) * elem_size + k]);
- }
- }
+ copyArrayToHost(send_offsets_id, ReadOnly);
+ copyArrayToHost(recv_offsets_id, ReadOnly);
+ copyArrayToHost(nsend_id, ReadOnly);
+ copyArrayToHost(nrecv_id, ReadOnly);
- PAIRS_DEBUG("\b]\n");
+ int nsend_all = 0;
+ int nrecv_all = 0;
+ for(int d = 0; d <= ndims; d++) {
+ nsend_all += nsend[d * 2 + 0];
+ nsend_all += nsend[d * 2 + 1];
+ nrecv_all += nrecv[d * 2 + 0];
+ nrecv_all += nrecv[d * 2 + 1];
}
- // Receive buffer debug
- for(int i = 0; i < 2; i++) {
- int nrec = nrecv[dim * 2 + i];
-
- PAIRS_DEBUG("recv_buf=[");
- for(int j = 0; j < MIN(elems_to_print, nrec); j++) {
- for(int k = 0; k < elem_size; k++) {
- PAIRS_DEBUG("%f,", recv_buf[(recv_offsets[dim * 2 + i] + j) * elem_size + k]);
- }
- }
-
- if(elems_to_print * 2 < nrec) {
- PAIRS_DEBUG("\b ... ");
- }
+ copyArrayToHost(send_buf_id, Ignore, nsend_all * elem_size * sizeof(real_t));
- for(int j = MAX(elems_to_print, nrec - elems_to_print); j < nrec; j++) {
- for(int k = 0; k < elem_size; k++) {
- PAIRS_DEBUG("%f,", recv_buf[(recv_offsets[dim * 2 + i] + j) * elem_size + k]);
- }
- }
+ array_flags->setHostFlag(recv_buf_id);
+ array_flags->clearDeviceFlag(recv_buf_id);
+ this->getDomainPartitioner()->communicateAllData(
+ ndims, elem_size, send_buf, send_offsets, nsend, recv_buf, recv_offsets, nrecv);
- PAIRS_DEBUG("\b]\n");
- }
- */
+ copyArrayToDevice(recv_buf_id, Ignore, nrecv_all * elem_size * sizeof(real_t));
}
void PairsSimulation::fillCommunicationArrays(int *neighbor_ranks, int *pbc, real_t *subdom) {
diff --git a/runtime/pairs.hpp b/runtime/pairs.hpp
index ba987e2..0cfbefb 100644
--- a/runtime/pairs.hpp
+++ b/runtime/pairs.hpp
@@ -229,11 +229,17 @@ public:
Regular6DStencil *getDomainPartitioner() { return dom_part; }
void communicateSizes(int dim, const int *send_sizes, int *recv_sizes);
+
void communicateData(
int dim, int elem_size,
const real_t *send_buf, const int *send_offsets, const int *nsend,
real_t *recv_buf, const int *recv_offsets, const int *nrecv);
+ void communicateAllData(
+ int ndims, int elem_size,
+ const real_t *send_buf, const int *send_offsets, const int *nsend,
+ real_t *recv_buf, const int *recv_offsets, const int *nrecv);
+
void fillCommunicationArrays(int neighbor_ranks[], int pbc[], real_t subdom[]);
// Device functions
diff --git a/src/pairs/sim/comm.py b/src/pairs/sim/comm.py
index 242c887..3f552ed 100644
--- a/src/pairs/sim/comm.py
+++ b/src/pairs/sim/comm.py
@@ -41,18 +41,31 @@ class Comm:
def synchronize(self):
# Every property that is not constant across timesteps and have neighbor accesses during any
# interaction kernel (i.e. property[j] in force calculation kernel)
- prop_list = [self.sim.property(p) for p in ['position', 'linear_velocity', 'angular_velocity'] if self.sim.property(p) is not None]
- for step in range(self.dom_part.number_of_steps()):
- PackGhostParticles(self, step, prop_list)
- CommunicateData(self, step, prop_list)
- UnpackGhostParticles(self, step, prop_list)
+ prop_names = ['position', 'linear_velocity', 'angular_velocity']
+ prop_list = [ self.sim.property(p) for p in prop_names if self.sim.property(p) is not None]
+
+ PackAllGhostParticles(self, prop_list)
+ CommunicateAllData(self, prop_list)
+ UnpackAllGhostParticles(self, prop_list)
@pairs_inline
def borders(self):
# Every property that has neighbor accesses during any interaction kernel (i.e. property[j]
# exists in any force calculation kernel)
- # We ignore normal because there should be no halfspace ghosts
- prop_list = [self.sim.property(p) for p in ['uid', 'type', 'mass', 'radius', 'position', 'linear_velocity', 'angular_velocity', 'shape'] if self.sim.property(p) is not None]
+ # We ignore normal because there should be no ghost half-spaces
+ prop_names = [
+ 'uid',
+ 'type',
+ 'mass',
+ 'radius',
+ 'position',
+ 'linear_velocity',
+ 'angular_velocity',
+ 'shape'
+ ]
+
+ prop_list = [self.sim.property(p) for p in prop_names if self.sim.property(p) is not None]
+
Assign(self.sim, self.nsend_all, 0)
Assign(self.sim, self.sim.nghost, 0)
@@ -75,12 +88,15 @@ class Comm:
PackGhostParticles(self, step, prop_list)
CommunicateData(self, step, prop_list)
UnpackGhostParticles(self, step, prop_list)
- Assign(self.sim, self.sim.nghost, self.sim.nghost + sum([self.nrecv[j] for j in self.dom_part.step_indexes(step)]))
+
+ step_nrecv = sum([self.nrecv[j] for j in self.dom_part.step_indexes(step)])
+ Assign(self.sim, self.sim.nghost, self.sim.nghost + step_nrecv)
@pairs_inline
def exchange(self):
# Every property except volatiles
prop_list = self.sim.properties.non_volatiles()
+
for step in range(self.dom_part.number_of_steps()):
Assign(self.sim, self.nsend_all, 0)
Assign(self.sim, self.sim.nghost, 0)
@@ -130,9 +146,31 @@ class CommunicateData(Lowerable):
@pairs_inline
def lower(self):
elem_size = sum([Types.number_of_elements(self.sim, p.type()) for p in self.prop_list])
- Call_Void(self.sim, "pairs->communicateData", [self.step, elem_size,
- self.comm.send_buffer, self.comm.send_offsets, self.comm.nsend,
- self.comm.recv_buffer, self.comm.recv_offsets, self.comm.nrecv])
+
+ Call_Void(self.sim,
+ "pairs->communicateData",
+ [self.step, elem_size,
+ self.comm.send_buffer, self.comm.send_offsets, self.comm.nsend,
+ self.comm.recv_buffer, self.comm.recv_offsets, self.comm.nrecv])
+
+
+class CommunicateAllData(Lowerable):
+ def __init__(self, comm, prop_list):
+ super().__init__(comm.sim)
+ self.comm = comm
+ self.prop_list = prop_list
+ self.sim.add_statement(self)
+
+ @pairs_inline
+ def lower(self):
+ elem_size = sum([Types.number_of_elements(self.sim, p.type()) for p in self.prop_list])
+
+ Call_Void(
+ self.sim,
+ "pairs->communicateAllData",
+ [self.comm.dom_part.number_of_steps(), elem_size,
+ self.comm.send_buffer, self.comm.send_offsets, self.comm.nsend,
+ self.comm.recv_buffer, self.comm.recv_offsets, self.comm.nrecv])
class DetermineGhostParticles(Lowerable):
@@ -281,6 +319,82 @@ class UnpackGhostParticles(Lowerable):
p_offset += 1
+class PackAllGhostParticles(Lowerable):
+ def __init__(self, comm, prop_list):
+ super().__init__(comm.sim)
+ self.comm = comm
+ self.prop_list = prop_list
+ self.sim.add_statement(self)
+
+ def get_elems_per_particle(self):
+ return sum([Types.number_of_elements(self.sim, p.type()) for p in self.prop_list])
+
+ @pairs_device_block
+ def lower(self):
+ send_buffer = self.comm.send_buffer
+ send_buffer.set_stride(1, self.get_elems_per_particle())
+ send_map = self.comm.send_map
+ send_mult = self.comm.send_mult
+ self.sim.module_name(f"pack_all_ghost_particles" + "_".join([str(p.id()) for p in self.prop_list]))
+
+ for i in For(self.sim, 0, self.comm.nsend_all):
+ p_offset = 0
+ m = send_map[i]
+ for p in self.prop_list:
+ if not Types.is_scalar(p.type()):
+ nelems = Types.number_of_elements(self.sim, p.type())
+ for e in range(nelems):
+ src = p[m][e]
+ if p == self.sim.position():
+ src += send_mult[i][e] * self.sim.grid.length(e)
+
+ Assign(self.sim, send_buffer[i][p_offset + e], src)
+
+ p_offset += nelems
+
+ else:
+ cast_fn = lambda x: Cast(self.sim, x, Types.Real) if p.type() != Types.Real else x
+ Assign(self.sim, send_buffer[i][p_offset], cast_fn(p[m]))
+ p_offset += 1
+
+
+class UnpackAllGhostParticles(Lowerable):
+ def __init__(self, comm, prop_list):
+ super().__init__(comm.sim)
+ self.comm = comm
+ self.prop_list = prop_list
+ self.sim.add_statement(self)
+
+ def get_elems_per_particle(self):
+ return sum([Types.number_of_elements(self.sim, p.type()) for p in self.prop_list])
+
+ @pairs_device_block
+ def lower(self):
+ nlocal = self.sim.nlocal
+ dom_part = self.comm.dom_part
+ recv_buffer = self.comm.recv_buffer
+ recv_buffer.set_stride(1, self.get_elems_per_particle())
+ self.sim.module_name(f"unpack_all_ghost_particles" + "_".join([str(p.id()) for p in self.prop_list]))
+
+ nrecv_size = sum([len(dom_part.step_indexes(s)) for s in range(dom_part.number_of_steps())])
+ nrecv_all = sum([self.comm.nrecv[j] for j in range(nrecv_size)])
+
+ for i in For(self.sim, 0, nrecv_all):
+ p_offset = 0
+ for p in self.prop_list:
+ if not Types.is_scalar(p.type()):
+ nelems = Types.number_of_elements(self.sim, p.type())
+ for e in range(nelems):
+ Assign(self.sim, p[nlocal + i][e], recv_buffer[i][p_offset + e])
+
+ p_offset += nelems
+
+ else:
+ cast_fn = lambda x: Cast(self.sim, x, p.type()) if p.type() != Types.Real else x
+ Assign(self.sim, p[nlocal + i], cast_fn(recv_buffer[i][p_offset]))
+ p_offset += 1
+
+
class RemoveExchangedParticles_part1(Lowerable):
def __init__(self, comm):
super().__init__(comm.sim)
--
GitLab