Initial commit: cuGenOpt GPU optimization solver

python/cugenopt/jit_template.cu

@@ -0,0 +1,204 @@
+/**
+ * JIT template — auto-generated by cugenopt.solve_custom()
+ *
+ * Placeholders (replaced by Python at runtime):
+ *   {{D1}}, {{D2}}             — Solution dimensions
+ *   {{ENCODING}}               — EncodingType enum value
+ *   {{ROW_MODE}}               — RowMode enum value
+ *   {{DIM1}}, {{DIM2_DEFAULT}}, {{TOTAL_ELEMENTS}}
+ *   {{VALUE_LOWER}}, {{VALUE_UPPER}}
+ *   {{CROSS_ROW_PROB}}, {{PERM_REPEAT_COUNT}}
+ *   {{OBJ_DEFS}}               — ObjDef array initializer
+ *   {{DATA_FIELDS}}            — struct field declarations
+ *   {{COMPUTE_OBJ}}            — user's compute_obj body
+ *   {{COMPUTE_PENALTY}}        — user's compute_penalty body
+ *   {{SHARED_MEM_EXPR}}        — shared_mem_bytes() return expression
+ *   {{LOAD_SHARED_BODY}}       — load_shared() body
+ *   {{DESTROY_BODY}}           — destroy() body
+ *   {{DATA_LOAD_BODY}}         — main() data loading code
+ *   {{SOLVER_CONFIG}}          — SolverConfig field assignments
+ *   {{NUM_OBJ}}                — number of objectives
+ *   {{CUSTOM_OP_DEFINES}}      — #define CUGENOPT_HAS_CUSTOM_OPS (or empty)
+ *   {{CUSTOM_OP_SWITCH}}       — switch cases for execute_custom_op (or empty)
+ *   {{CUSTOM_OP_REGISTRY}}     — add() calls for register_custom_operators (or empty)
+ */
+
+{{CUSTOM_OP_DEFINES}}
+
+#include "types.cuh"
+#include "cuda_utils.cuh"
+
+struct CustomProblem;
+
+#ifdef CUGENOPT_HAS_CUSTOM_OPS
+template<typename Sol>
+__device__ inline bool execute_custom_op(int seq_id, Sol& sol, int dim1,
+                                          EncodingType encoding, curandState* rng,
+                                          int val_lb, int val_ub,
+                                          const void* prob_data);
+#endif
+
+#include "solver.cuh"
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+
+struct CustomProblem : ProblemBase<CustomProblem, {{D1}}, {{D2}}> {
+    {{DATA_FIELDS}}
+    int _n;
+
+    static constexpr ObjDef OBJ_DEFS[] = {
+        {{OBJ_DEFS}}
+    };
+
+    __device__ float compute_obj(int idx, const Sol& sol) const {
+        {{COMPUTE_OBJ}}
+    }
+
+    __device__ float compute_penalty(const Sol& sol) const {
+        {{COMPUTE_PENALTY}}
+    }
+
+    ProblemConfig config() const {
+        ProblemConfig cfg;
+        cfg.encoding = {{ENCODING}};
+        cfg.dim1 = {{DIM1}};
+        cfg.dim2_default = {{DIM2_DEFAULT}};
+        cfg.row_mode = {{ROW_MODE}};
+        cfg.total_elements = {{TOTAL_ELEMENTS}};
+        cfg.cross_row_prob = {{CROSS_ROW_PROB}};
+        cfg.perm_repeat_count = {{PERM_REPEAT_COUNT}};
+        cfg.value_lower_bound = {{VALUE_LOWER}};
+        cfg.value_upper_bound = {{VALUE_UPPER}};
+        fill_obj_config(cfg);
+        return cfg;
+    }
+
+    size_t shared_mem_bytes() const {
+        {{SHARED_MEM_EXPR}}
+    }
+
+    __device__ void load_shared(char* smem, int tid, int bsz) {
+        {{LOAD_SHARED_BODY}}
+    }
+
+    void destroy() {
+        {{DESTROY_BODY}}
+    }
+};
+
+#ifdef CUGENOPT_HAS_CUSTOM_OPS
+template<typename Sol>
+__device__ inline bool execute_custom_op(int seq_id, Sol& sol, int dim1,
+                                          EncodingType encoding, curandState* rng,
+                                          int val_lb, int val_ub,
+                                          const void* prob_data) {
+    const CustomProblem* prob = static_cast<const CustomProblem*>(prob_data);
+    switch (seq_id) {
+{{CUSTOM_OP_SWITCH}}
+    default:
+        return false;
+    }
+}
+#endif
+
+#ifdef CUGENOPT_HAS_CUSTOM_OPS
+inline void register_custom_operators(SeqRegistry& reg) {
+    auto add = [&](int id, float w, float cap) {
+        if (reg.count >= MAX_SEQ) return;
+        reg.ids[reg.count] = id;
+        reg.weights[reg.count] = w;
+        reg.max_w[reg.count] = cap;
+        reg.count++;
+    };
+{{CUSTOM_OP_REGISTRY}}
+    float sum = 0.0f;
+    for (int i = 0; i < reg.count; i++) sum += reg.weights[i];
+    if (sum > 0.0f) {
+        for (int i = 0; i < reg.count; i++) reg.weights[i] /= sum;
+    }
+}
+#endif
+
+static float* read_binary_floats(const char* path, int count) {
+    FILE* f = fopen(path, "rb");
+    if (!f) { fprintf(stderr, "Cannot open %s\n", path); exit(1); }
+    float* buf = new float[count];
+    size_t r = fread(buf, sizeof(float), count, f);
+    fclose(f);
+    if ((int)r != count) { fprintf(stderr, "Short read %s: %d/%d\n", path, (int)r, count); exit(1); }
+    return buf;
+}
+
+static int* read_binary_ints(const char* path, int count) {
+    FILE* f = fopen(path, "rb");
+    if (!f) { fprintf(stderr, "Cannot open %s\n", path); exit(1); }
+    int* buf = new int[count];
+    size_t r = fread(buf, sizeof(int), count, f);
+    fclose(f);
+    if ((int)r != count) { fprintf(stderr, "Short read %s: %d/%d\n", path, (int)r, count); exit(1); }
+    return buf;
+}
+
+int main(int argc, char** argv) {
+    if (argc < 2) {
+        fprintf(stderr, "Usage: %s <data_dir>\n", argv[0]);
+        return 1;
+    }
+    const char* data_dir = argv[1];
+    char path[512];
+
+    // Read problem size
+    snprintf(path, sizeof(path), "%s/n.bin", data_dir);
+    int n;
+    { FILE* f = fopen(path, "rb"); fread(&n, sizeof(int), 1, f); fclose(f); }
+
+    // Read data arrays and upload to GPU
+    CustomProblem prob;
+    prob._n = n;
+    {{DATA_LOAD_BODY}}
+
+    // Configure solver
+    SolverConfig cfg;
+    {{SOLVER_CONFIG}}
+
+    // Solve
+#ifdef CUGENOPT_HAS_CUSTOM_OPS
+    auto result = solve(prob, cfg, nullptr, 0, register_custom_operators);
+#else
+    auto result = solve(prob, cfg);
+#endif
+
+    // Output JSON result
+    auto& best = result.best_solution;
+    printf("{\"objective\":%.10g", best.objectives[0]);
+    printf(",\"penalty\":%.10g", best.penalty);
+    printf(",\"elapsed_ms\":%.4f", result.elapsed_ms);
+    printf(",\"generations\":%d", result.generations);
+    printf(",\"stop_reason\":\"%s\"",
+           result.stop_reason == StopReason::TimeLimit ? "time_limit" :
+           result.stop_reason == StopReason::Stagnation ? "stagnation" : "max_gen");
+
+    printf(",\"objectives\":[");
+    for (int i = 0; i < {{NUM_OBJ}}; i++) {
+        if (i > 0) printf(",");
+        printf("%.10g", best.objectives[i]);
+    }
+    printf("]");
+
+    printf(",\"solution\":[");
+    for (int r = 0; r < {{DIM1}}; r++) {
+        if (r > 0) printf(",");
+        printf("[");
+        int len = best.dim2_sizes[r];
+        for (int c = 0; c < len; c++) {
+            if (c > 0) printf(",");
+            printf("%d", best.data[r][c]);
+        }
+        printf("]");
+    }
+    printf("]}\n");
+
+    prob.destroy();
+    return 0;
+}