Rework AI neighborlist slop from scratch

kernels/neighbor_list.cu

@@ -1,194 +1 @@
 #include "neighbor_list.cuh"
-
-/**
- * Step 1: Assign particles to cells
- */
-__global__ void assign_particles_to_cells_kernel(const float4 *positions,
-                                                 int *particle_cells,
-                                                 const CellList cell_list,
-                                                 size_t n_particles) {
-  size_t i = get_thread_id();
-  if (i >= n_particles)
-    return;
-
-  float4 pos = positions[i];
-  float3 pos3 = make_float3(pos.x, pos.y, pos.z);
-
-  int3 cell_coords = cell_list.get_cell_coords(pos3);
-
-  // Clamp to valid range (handle edge cases)
-  cell_coords.x = max(0, min(cell_coords.x, cell_list.grid_size.x - 1));
-  cell_coords.y = max(0, min(cell_coords.y, cell_list.grid_size.y - 1));
-  cell_coords.z = max(0, min(cell_coords.z, cell_list.grid_size.z - 1));
-
-  particle_cells[i] =
-      cell_list.get_cell_index(cell_coords.x, cell_coords.y, cell_coords.z);
-}
-
-/**
- * Step 2: Find cell boundaries after sorting
- */
-__global__ void find_cell_boundaries_kernel(const int *sorted_particle_cells,
-                                            int *cell_starts, int *cell_ends,
-                                            size_t n_particles,
-                                            size_t total_cells) {
-  size_t i = get_thread_id();
-  if (i >= n_particles)
-    return;
-
-  int cell = sorted_particle_cells[i];
-
-  // Check if this is the start of a new cell
-  if (i == 0 || sorted_particle_cells[i - 1] != cell) {
-    cell_starts[cell] = i;
-  }
-
-  // Check if this is the end of a cell
-  if (i == n_particles - 1 || sorted_particle_cells[i + 1] != cell) {
-    cell_ends[cell] = i + 1;
-  }
-}
-
-/**
- * Step 3: Build actual neighbor lists using cell lists
- */
-__global__ void
-build_neighbor_lists_kernel(const float4 *positions, const CellList cell_list,
-                            NeighborList neighbor_list, float cutoff_squared,
-                            const float3 *box_lengths, size_t n_particles) {
-  size_t i = get_thread_id();
-  if (i >= n_particles)
-    return;
-
-  float4 pos_i = positions[i];
-  float3 my_pos = make_float3(pos_i.x, pos_i.y, pos_i.z);
-
-  int3 my_cell = cell_list.get_cell_coords(my_pos);
-  int neighbor_count = 0;
-  int max_neighbors = neighbor_list.max_neighbors;
-  int *my_neighbors = neighbor_list.get_neighbors(i);
-
-  // Search neighboring cells (3x3x3 = 27 cells including self)
-  for (int dz = -1; dz <= 1; dz++) {
-    for (int dy = -1; dy <= 1; dy++) {
-      for (int dx = -1; dx <= 1; dx++) {
-
-        int3 neighbor_cell =
-            make_int3(my_cell.x + dx, my_cell.y + dy, my_cell.z + dz);
-
-        // Check bounds
-        if (neighbor_cell.x < 0 || neighbor_cell.x >= cell_list.grid_size.x ||
-            neighbor_cell.y < 0 || neighbor_cell.y >= cell_list.grid_size.y ||
-            neighbor_cell.z < 0 || neighbor_cell.z >= cell_list.grid_size.z) {
-          continue;
-        }
-
-        int cell_idx = cell_list.get_cell_index(
-            neighbor_cell.x, neighbor_cell.y, neighbor_cell.z);
-        int start = cell_list.cell_starts[cell_idx];
-        int end = cell_list.cell_ends[cell_idx];
-
-        // Check all particles in this cell
-        for (int idx = start; idx < end; idx++) {
-          int j = cell_list.sorted_particles[idx];
-
-          if (i >= j)
-            continue; // Avoid double counting and self-interaction
-
-          float4 pos_j = positions[j];
-          float3 other_pos = make_float3(pos_j.x, pos_j.y, pos_j.z);
-
-          // Calculate distance with periodic boundary conditions
-          float dx = my_pos.x - other_pos.x;
-          float dy = my_pos.y - other_pos.y;
-          float dz = my_pos.z - other_pos.z;
-
-          // Apply PBC
-          dx -= box_lengths->x * roundf(dx / box_lengths->x);
-          dy -= box_lengths->y * roundf(dy / box_lengths->y);
-          dz -= box_lengths->z * roundf(dz / box_lengths->z);
-
-          float r_squared = dx * dx + dy * dy + dz * dz;
-
-          if (r_squared <= cutoff_squared && neighbor_count < max_neighbors) {
-            my_neighbors[neighbor_count] = j;
-            neighbor_count++;
-          }
-        }
-      }
-    }
-  }
-
-  neighbor_list.neighbor_counts[i] = neighbor_count;
-}
-
-// =============================================================================
-// HOST FUNCTIONS
-// =============================================================================
-
-void build_cell_list(const float4 *positions, CellList &cell_list,
-                     const float3 *box_lengths) {
-
-  // Step 1: Assign particles to cells
-  auto config = get_launch_config(cell_list.n_particles);
-
-  assign_particles_to_cells_kernel<<<config.blocks, config.threads>>>(
-      positions, cell_list.particle_cells, cell_list, cell_list.n_particles);
-
-  // Step 2: Sort particles by cell index using Thrust
-  thrust::device_ptr<int> particle_cells_ptr(cell_list.particle_cells);
-  thrust::device_ptr<int> sorted_particles_ptr(cell_list.sorted_particles);
-
-  // Initialize particle indices 0, 1, 2, ...
-  thrust::sequence(sorted_particles_ptr,
-                   sorted_particles_ptr + cell_list.n_particles);
-
-  // Sort particle indices by their cell assignments
-  thrust::sort_by_key(particle_cells_ptr,
-                      particle_cells_ptr + cell_list.n_particles,
-                      sorted_particles_ptr);
-
-  // Step 3: Initialize cell boundaries to -1
-  cudaMemset(cell_list.cell_starts, -1, cell_list.total_cells * sizeof(int));
-  cudaMemset(cell_list.cell_ends, -1, cell_list.total_cells * sizeof(int));
-
-  // Step 4: Find cell boundaries
-  config = get_launch_config(cell_list.n_particles);
-  find_cell_boundaries_kernel<<<config.blocks, config.threads>>>(
-      cell_list.particle_cells, cell_list.cell_starts, cell_list.cell_ends,
-      cell_list.n_particles, cell_list.total_cells);
-}
-
-void build_neighbor_list(const float4 *positions, const CellList &cell_list,
-                         NeighborList &neighbor_list, float cutoff_squared,
-                         const float3 *box_lengths) {
-
-  // Initialize neighbor counts to 0
-  cudaMemset(neighbor_list.neighbor_counts, 0,
-             neighbor_list.n_particles * sizeof(int));
-
-  auto config = get_launch_config(neighbor_list.n_particles);
-
-  build_neighbor_lists_kernel<<<config.blocks, config.threads>>>(
-      positions, cell_list, neighbor_list, cutoff_squared, box_lengths,
-      neighbor_list.n_particles);
-}
-
-void build_neighbor_list_optimized(const float4 *positions, size_t n_particles,
-                                   const float3 *box_lengths, float cutoff,
-                                   NeighborList &neighbor_list) {
-
-  // Get box size for cell list construction
-  float3 box_size;
-  cudaMemcpy(&box_size, box_lengths, sizeof(float3), cudaMemcpyDeviceToHost);
-
-  // Create cell list
-  CellList cell_list(n_particles, box_size, cutoff);
-
-  // Build cell list
-  build_cell_list(positions, cell_list, box_lengths);
-
-  // Build neighbor list using cell list
-  build_neighbor_list(positions, cell_list, neighbor_list, cutoff * cutoff,
-                      box_lengths);
-}

kernels/neighbor_list.cuh

@@ -1,68 +1,21 @@
 #ifndef NEIGHBOR_LIST_CUH
 #define NEIGHBOR_LIST_CUH
 
-#include "kernel_config.cuh" // From previous artifact
+#include "box.hpp"
+#include "kernel_config.cuh"
+#include "utils.cuh"
 #include <cuda_runtime.h>
 #include <thrust/device_vector.h>
+#include <thrust/pair.h>
 #include <thrust/sequence.h>
 #include <thrust/sort.h>
 
-/**
- * Simple kernel to initialize neighbor offsets
- */
-__global__ void init_neighbor_offsets(int *offsets, size_t n_particles,
-                                      int max_neighbors) {
-  size_t i = get_thread_id();
-  if (i <= n_particles) { // Note: <= because we need n_particles + 1 elements
-    offsets[i] = i * max_neighbors;
-  }
-}
-
-/**
- * Neighbor list data structure
- * Uses a compact format similar to CSR sparse matrices
- */
-struct NeighborList {
-  int *neighbor_offsets; // Size: n_particles + 1, offset into neighbor_indices
-  int *neighbor_indices; // Size: total_neighbors, actual neighbor particle IDs
-  int *neighbor_counts;  // Size: n_particles, number of neighbors per particle
-  size_t max_neighbors;  // Maximum neighbors allocated per particle
-  size_t n_particles;
-
-  // Constructor
-  NeighborList(size_t n_particles, size_t max_neighbors_per_particle)
-      : max_neighbors(max_neighbors_per_particle), n_particles(n_particles) {
-
-    cudaMalloc(&neighbor_offsets, (n_particles + 1) * sizeof(int));
-    cudaMalloc(&neighbor_indices, n_particles * max_neighbors * sizeof(int));
-    cudaMalloc(&neighbor_counts, n_particles * sizeof(int));
-
-    // Initialize offsets
-    auto kernel_config = get_launch_config(n_particles);
-    init_neighbor_offsets<<<kernel_config.blocks, kernel_config.threads>>>(
-        neighbor_offsets, n_particles, max_neighbors);
-  }
-
-  ~NeighborList() {
-    cudaFree(neighbor_offsets);
-    cudaFree(neighbor_indices);
-    cudaFree(neighbor_counts);
-  }
-
-  // Get neighbors for particle i (device function)
-  __device__ int *get_neighbors(int i) const {
-    return &neighbor_indices[neighbor_offsets[i]];
-  }
-
-  __device__ int get_neighbor_count(int i) const { return neighbor_counts[i]; }
-};
-
 /**
  * Cell list structure for spatial hashing
  */
 struct CellList {
   int *cell_starts;      // Size: total_cells, start index in sorted_particles
-  int *cell_ends;        // Size: total_cells, end index in sorted_particles
+  int *cell_count;       // Size: total_cells, end index in sorted_particles
   int *sorted_particles; // Size: n_particles, particle IDs sorted by cell
   int *particle_cells; // Size: n_particles, which cell each particle belongs to
 
@@ -72,31 +25,26 @@ struct CellList {
   size_t total_cells;
   size_t n_particles;
 
-  CellList(size_t n_particles, float3 box_size, float cutoff)
+  CellList(size_t n_particles, Box &box, float cutoff)
       : n_particles(n_particles) {
 
-    // Calculate grid dimensions (cell size slightly larger than cutoff)
+    box_min.x = box.xlo;
-    float cell_margin = 1.001f; // Small safety margin
+    box_min.y = box.ylo;
-    cell_size = make_float3(cutoff * cell_margin, cutoff * cell_margin,
+    box_min.z = box.zlo;
-                            cutoff * cell_margin);
 
-    grid_size.x = (int)ceilf(box_size.x / cell_size.x);
+    auto [grid_size, cell_size] = calc_grid_and_cell_size(box, cutoff);
-    grid_size.y = (int)ceilf(box_size.y / cell_size.y);
-    grid_size.z = (int)ceilf(box_size.z / cell_size.z);
 
-    total_cells = (size_t)grid_size.x * grid_size.y * grid_size.z;
+    total_cells = grid_size.x * grid_size.y * grid_size.z;
-    box_min = make_float3(0.0f, 0.0f, 0.0f); // Assume box starts at origin
 
-    // Allocate memory
     cudaMalloc(&cell_starts, total_cells * sizeof(int));
-    cudaMalloc(&cell_ends, total_cells * sizeof(int));
+    cudaMalloc(&cell_count, total_cells * sizeof(int));
     cudaMalloc(&sorted_particles, n_particles * sizeof(int));
     cudaMalloc(&particle_cells, n_particles * sizeof(int));
   }
 
   ~CellList() {
     cudaFree(cell_starts);
-    cudaFree(cell_ends);
+    cudaFree(cell_count);
     cudaFree(sorted_particles);
     cudaFree(particle_cells);
   }
@@ -106,6 +54,20 @@ struct CellList {
     return z * grid_size.x * grid_size.y + y * grid_size.x + x;
   }
 
+  std::pair<int3, float3> calc_grid_and_cell_size(Box &box, float cutoff) {
+    int3 grid_size = {utils::max((int)(box.xhi - box.xlo) / cutoff, 1),
+                      utils::max((int)(box.yhi - box.ylo) / cutoff, 1),
+                      utils::max((int)(box.zhi - box.zlo) / cutoff, 1)};
+
+    float3 cell_size = {
+        (box.xhi - box.xlo) / grid_size.x,
+        (box.yhi - box.ylo) / grid_size.y,
+        (box.zhi - box.zlo) / grid_size.z,
+    };
+
+    return std::make_pair(grid_size, cell_size);
+  }
+
   // Get cell coordinates from position
   __device__ int3 get_cell_coords(float3 pos) const {
     return make_int3((int)((pos.x - box_min.x) / cell_size.x),
@@ -114,19 +76,4 @@ struct CellList {
   }
 };
 
-// Forward declarations
-void build_cell_list(const float4 *positions, CellList &cell_list,
-                     const float3 *box_lengths);
-
-void build_neighbor_list(const float4 *positions, const CellList &cell_list,
-                         NeighborList &neighbor_list, float cutoff_squared,
-                         const float3 *box_lengths);
-
-/**
- * High-level interface - builds complete neighbor list
- */
-void build_neighbor_list_optimized(const float4 *positions, size_t n_particles,
-                                   const float3 *box_lengths, float cutoff,
-                                   NeighborList &neighbor_list);
-
 #endif

kernels/utils.cuh

@@ -0,0 +1,8 @@
+#ifndef UTILS_CUH
+#define UTILS_CUH
+
+namespace utils {
+__device__ __host__ inline int max(int a, int b) { return (a > b) ? a : b; }
+} // namespace utils
+
+#endif