Rewrite the force calculations to fix memory issues

kernels/forces.cuh

@@ -1,19 +1,80 @@
 #ifndef FORCES_CUH
 #define FORCES_CUH
-
-#include "pair_potentials.cuh"
+#include "potentials/pair_potentials.cuh"
 #include "precision.hpp"
+#include <cstdio>
+#include <type_traits>
+#include <variant>
+#include <vector>
+
 namespace CAC {
-/**
- * Calculate forces and energies using CUDA for acceleration
- * This code currently only accepts a single PairPotential object and does an
- * n^2 force calculation. Future improvements will:
- *   - Allow for neighbor listing
- *   - Allow for overlaid force calculations
- */
+
+inline void reset_forces_and_energies(int n_particles, real *forces,
+                                      real *energies) {
+  cudaMemset(forces, 0, n_particles * sizeof(real) * 3);
+  cudaMemset(energies, 0, n_particles * sizeof(real));
+}
+
+template <typename PotentialType>
 __global__ void calc_forces_and_energies(real *xs, real *forces, real *energies,
-                                         int n_particles, real *box_bd,
-                                         PairPotential &potential);
+                                         int n_particles, real *box_len,
+                                         PotentialType potential) {
+  int i = blockIdx.x * blockDim.x + threadIdx.x;
+
+  if (i == 0) {
+    printf("n_particles: %d\n", n_particles);
+    printf("box_len: %f %f %f\n", box_len[0], box_len[1], box_len[2]);
+  }
+
+  if (i < n_particles) {
+    printf("Thread %d, Block %d\n", threadIdx.x, blockIdx.x);
+    real xi = xs[3 * i];
+    real yi = xs[3 * i + 1];
+    real zi = xs[3 * i + 2];
+
+    for (int j = 0; j < n_particles; j++) {
+      if (i != j) {
+        real xj = xs[3 * j];
+        real yj = xs[3 * j + 1];
+        real zj = xs[3 * j + 2];
+
+        real dx = xi - xj;
+        real dy = yi - yj;
+        real dz = zi - zj;
+
+        // Apply periodic boundary conditions
+        dx -= box_len[0] * round(dx / box_len[0]);
+        dy -= box_len[1] * round(dy / box_len[1]);
+        dz -= box_len[2] * round(dz / box_len[2]);
+
+        ForceAndEnergy sol = potential.calc_force_and_energy({dx, dy, dz});
+        forces[3 * i] += sol.force.x;
+        forces[3 * i + 1] += sol.force.y;
+        forces[3 * i + 2] += sol.force.z;
+        energies[i] += sol.energy;
+      }
+    }
+  }
+}
+
+inline void launch_force_kernels(real *xs, real *forces, real *energies,
+                                 int n_particles, real *box_len,
+                                 std::vector<PairPotentials> potentials,
+                                 int grid_size, int block_size) {
+
+  reset_forces_and_energies(n_particles, forces, energies);
+
+  for (const auto &potential : potentials) {
+    std::visit(
+        [&](const auto &potential) {
+          using PotentialType = std::decay_t<decltype(potential)>;
+          calc_forces_and_energies<PotentialType><<<grid_size, block_size>>>(
+              xs, forces, energies, n_particles, box_len, potential);
+        },
+        potential);
+    cudaDeviceSynchronize();
+  }
+}
 } // namespace CAC
 
 #endif