Skip to content

Optimize the VC-SGC MC trial loop in the NEP path #1455

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

Draft
dingxu2016 wants to merge 2 commits into
base: master
Choose a base branch
from
+765 −41
Draft

Optimize the VC-SGC MC trial loop in the NEP path #1455

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
55f1ad0
Fuse dual-state VC-SGC NEP evaluation with L_max dispatch
dingxu2016 Apr 18, 2026
96bd2a7
Reuse global shell neighbors for VC-SGC local inputs
dingxu2016 Apr 20, 2026
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
182 changes: 141 additions & 41 deletions src/mc/mc_ensemble_sgc.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,9 +27,11 @@ integration across phase boundaries, Phys. Rev. B 86, 134204 (2012).
------------------------------------------------------------------------------*/

#include "mc_ensemble_sgc.cuh"
#include "force/neighbor.cuh"
#include "utilities/gpu_macro.cuh"
#include <algorithm>
#include <cmath>
#include <map>
#include <cstring>

const std::map<std::string, double> MASS_TABLE{
{"H", 1.0080000000},
Expand Down Expand Up @@ -287,6 +289,74 @@ static __global__ void create_inputs_for_energy_calculator(
}
}

static __global__ void create_inputs_for_energy_calculator_from_global_shell(
const int N,
const int N_local,
const int global_shell_stride,
const int* atom_local,
const Box box,
const float rc_radial_square,
const float rc_angular_square,
const double* __restrict__ g_x,
const double* __restrict__ g_y,
const double* __restrict__ g_z,
const int* __restrict__ g_type_before,
const int* __restrict__ g_type_after,
const int* __restrict__ g_NN_global,
const int* __restrict__ g_NL_global,
int* g_NN_radial,
int* g_NN_angular,
int* g_t2_radial_before,
int* g_t2_radial_after,
int* g_t2_angular_before,
int* g_t2_angular_after,
float* g_x12_radial,
float* g_y12_radial,
float* g_z12_radial,
float* g_x12_angular,
float* g_y12_angular,
float* g_z12_angular)
{
int candidate = blockIdx.x * blockDim.x + threadIdx.x;
int total_candidates = N_local * global_shell_stride;
if (candidate >= total_candidates) {
return;
}

int k = candidate % N_local;
int slot = candidate / N_local;
int n1 = atom_local[k];
if (slot >= g_NN_global[n1]) {
return;
}

int n2 = g_NL_global[n1 + N * slot];
double x12 = g_x[n2] - g_x[n1];
double y12 = g_y[n2] - g_y[n1];
double z12 = g_z[n2] - g_z[n1];
apply_mic(box, x12, y12, z12);
float distance_square = float(x12 * x12 + y12 * y12 + z12 * z12);

if (distance_square < rc_radial_square) {
int count_radial = atomicAdd(&g_NN_radial[k], 1);
int index_radial = count_radial * N_local + k;
g_t2_radial_before[index_radial] = g_type_before[n2];
g_t2_radial_after[index_radial] = g_type_after[n2];
g_x12_radial[index_radial] = float(x12);
g_y12_radial[index_radial] = float(y12);
g_z12_radial[index_radial] = float(z12);
}
if (distance_square < rc_angular_square) {
int count_angular = atomicAdd(&g_NN_angular[k], 1);
int index_angular = count_angular * N_local + k;
g_t2_angular_before[index_angular] = g_type_before[n2];
g_t2_angular_after[index_angular] = g_type_after[n2];
g_x12_angular[index_angular] = float(x12);
g_y12_angular[index_angular] = float(y12);
g_z12_angular[index_angular] = float(z12);
}
}

// a kernel with a single thread <<<1, 1>>>
static __global__ void gpu_flip(
const int i,
Expand Down Expand Up @@ -326,7 +396,9 @@ void MC_Ensemble_SGC::compute(
int grouping_method,
int group_id)
{
if (check_if_small_box(nep_energy.paramb.rc_radial_max, box)) {
float required_shell_cutoff =
std::max(nep_energy.paramb.rc_radial_max, nep_energy.paramb.rc_angular_max);
if (check_if_small_box(required_shell_cutoff, box)) {
printf("Cannot use small box for MCMD.\n");
exit(1);
}
Expand All @@ -336,6 +408,13 @@ void MC_Ensemble_SGC::compute(
type_after.resize(atom.number_of_atoms);
}

Neighbor global_shell_neighbor;
int seed_neighbors = std::max(nep_energy.paramb.MN_radial, nep_energy.paramb.MN_angular);
global_shell_neighbor.initialize(required_shell_cutoff, atom.number_of_atoms, seed_neighbors);
global_shell_neighbor.find_neighbor_global(
required_shell_cutoff, box, atom.type, atom.position_per_atom);
int global_shell_stride = int(global_shell_neighbor.NL.size() / atom.number_of_atoms);

int group_size =
grouping_method >= 0 ? groups[grouping_method].cpu_size[group_id] : atom.number_of_atoms;
std::uniform_int_distribution<int> r1(0, group_size - 1);
Expand Down Expand Up @@ -392,60 +471,81 @@ void MC_Ensemble_SGC::compute(

CHECK(gpuMemset(NN_radial.data(), 0, sizeof(int) * NN_radial.size()));
CHECK(gpuMemset(NN_angular.data(), 0, sizeof(int) * NN_angular.size()));
create_inputs_for_energy_calculator<<<(atom.number_of_atoms - 1) / 64 + 1, 64>>>(
atom.number_of_atoms,
NN_ij_cpu,
NL_ij.data(),
box,
nep_energy.paramb.rc_radial_max * nep_energy.paramb.rc_radial_max,
nep_energy.paramb.rc_angular_max * nep_energy.paramb.rc_angular_max,
atom.position_per_atom.data(),
atom.position_per_atom.data() + atom.number_of_atoms,
atom.position_per_atom.data() + atom.number_of_atoms * 2,
type_before.data(),
type_after.data(),
NN_radial.data(),
NN_angular.data(),
t2_radial_before.data(),
t2_radial_after.data(),
t2_angular_before.data(),
t2_angular_after.data(),
x12_radial.data(),
y12_radial.data(),
z12_radial.data(),
x12_angular.data(),
y12_angular.data(),
z12_angular.data());
if (global_shell_stride > 0) {
int total_candidates = NN_ij_cpu * global_shell_stride;
create_inputs_for_energy_calculator_from_global_shell<<<
(total_candidates - 1) / 128 + 1, 128>>>(
atom.number_of_atoms,
NN_ij_cpu,
global_shell_stride,
NL_ij.data(),
box,
nep_energy.paramb.rc_radial_max * nep_energy.paramb.rc_radial_max,
nep_energy.paramb.rc_angular_max * nep_energy.paramb.rc_angular_max,
atom.position_per_atom.data(),
atom.position_per_atom.data() + atom.number_of_atoms,
atom.position_per_atom.data() + atom.number_of_atoms * 2,
type_before.data(),
type_after.data(),
global_shell_neighbor.NN.data(),
global_shell_neighbor.NL.data(),
NN_radial.data(),
NN_angular.data(),
t2_radial_before.data(),
t2_radial_after.data(),
t2_angular_before.data(),
t2_angular_after.data(),
x12_radial.data(),
y12_radial.data(),
z12_radial.data(),
x12_angular.data(),
y12_angular.data(),
z12_angular.data());
} else {
create_inputs_for_energy_calculator<<<(atom.number_of_atoms - 1) / 64 + 1, 64>>>(
atom.number_of_atoms,
NN_ij_cpu,
NL_ij.data(),
box,
nep_energy.paramb.rc_radial_max * nep_energy.paramb.rc_radial_max,
nep_energy.paramb.rc_angular_max * nep_energy.paramb.rc_angular_max,
atom.position_per_atom.data(),
atom.position_per_atom.data() + atom.number_of_atoms,
atom.position_per_atom.data() + atom.number_of_atoms * 2,
type_before.data(),
type_after.data(),
NN_radial.data(),
NN_angular.data(),
t2_radial_before.data(),
t2_radial_after.data(),
t2_angular_before.data(),
t2_angular_after.data(),
x12_radial.data(),
y12_radial.data(),
z12_radial.data(),
x12_angular.data(),
y12_angular.data(),
z12_angular.data());
}
GPU_CHECK_KERNEL

nep_energy.find_energy(
nep_energy.find_energy_dual(
NN_ij_cpu,
NN_radial.data(),
NN_angular.data(),
local_type_before.data(),
t2_radial_before.data(),
t2_angular_before.data(),
x12_radial.data(),
y12_radial.data(),
z12_radial.data(),
x12_angular.data(),
y12_angular.data(),
z12_angular.data(),
pe_before.data());

nep_energy.find_energy(
NN_ij_cpu,
NN_radial.data(),
NN_angular.data(),
local_type_after.data(),
t2_radial_before.data(),
t2_radial_after.data(),
t2_angular_before.data(),
t2_angular_after.data(),
x12_radial.data(),
y12_radial.data(),
z12_radial.data(),
x12_angular.data(),
y12_angular.data(),
z12_angular.data(),
pe_before.data(),
pe_after.data());

std::vector<float> pe_before_cpu(NN_ij_cpu);
Expand Down
Loading