Add GPU implementation of preconditioner - v2 (#31)

* New feat: GPU precondition

* New feat: synchronized A/At scaling

* Todo: infeasible is stucked

* New feat: preconditioner prints

* Bug fixed: objective & bound rescale

* feat: add logging for precondition time

* Bug fixed: numerical issue

* Clean code: clang-format and align naming style

* Clean code: rename variables

* Refactor: move finite-bound computation

* Numerical issue: back to old version

* Code refactor: CUB DeviceReduce for bound norm

* Bug fixed: typo

* Clean code: Limit CUB headers to device_reduce to reduce compile time

* Clean code

* Bug fixed: warm start solution

* Bug fixed: illegal memory access in bound

* Bug fixed: norm computation on scaled bound

* Revert "Bug fixed: norm computation on scaled bound"

This reverts commit d372a20.

* Bug fixed: unused sv params

* New feat: add verbose logging

* New ver: 0.2.4

Author: LucasBoTang

.gitignore

@@ -63,3 +63,7 @@ test/*
 /_b
 *.whl
 *.pyc
+
+
+*.txt
+*.sh

CMakeLists.txt

@@ -8,7 +8,7 @@ project(cupdlpx LANGUAGES C CXX CUDA)
 
 set(CUPDLPX_VERSION_MAJOR 0)
 set(CUPDLPX_VERSION_MINOR 2)
-set(CUPDLPX_VERSION_PATCH 3)
+set(CUPDLPX_VERSION_PATCH 4)
 
 set(CUPDLPX_VERSION "${CUPDLPX_VERSION_MAJOR}.${CUPDLPX_VERSION_MINOR}.${CUPDLPX_VERSION_PATCH}")
 add_compile_definitions(CUPDLPX_VERSION="${CUPDLPX_VERSION}")

internal/internal_types.h

@@ -28,7 +28,9 @@ typedef struct
 	int num_nonzeros;
 	int *row_ptr;
 	int *col_ind;
+	int *row_ind;
 	double *val;
+	int *transpose_map;
 } cu_sparse_matrix_csr_t;
 
 typedef struct
@@ -46,6 +48,7 @@ typedef struct
 	int num_blocks_primal;
 	int num_blocks_dual;
 	int num_blocks_primal_dual;
+	int num_blocks_nnz;
 	double objective_vector_norm;
 	double constraint_bound_norm;
 	double *constraint_lower_bound_finite_val;
@@ -129,7 +132,6 @@ typedef struct
 
 typedef struct
 {
-	lp_problem_t *scaled_problem;
 	double *con_rescale;
 	double *var_rescale;
 	double con_bound_rescale;

internal/preconditioner.h

@@ -25,7 +25,7 @@ extern "C"
 
     rescale_info_t *rescale_problem(
         const pdhg_parameters_t *params,
-        const lp_problem_t *original_problem);
+        pdhg_solver_state_t *state);
 
 #ifdef __cplusplus
 }

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "scikit_build_core.build"
 
 [project]
 name = "cupdlpx"
-version = "0.2.2"
+version = "0.2.4"
 description = "Python bindings for cuPDLPx (GPU-accelerated first-order LP solver)"
 readme = "README.md"
 license = { text = "Apache-2.0" }
@@ -25,7 +25,7 @@ wheel.packages = ["python/cupdlpx"]
 sdist.include = ["tests/**", "pyproject.toml", "README.md", "LICENSE"]
 
 [tool.scikit-build.cmake.define]
-CMAKE_CUDA_ARCHITECTURES = "60;61;70;75;80;86;89;90;90-virtual"
+CMAKE_CUDA_ARCHITECTURES = "all"
 CMAKE_CUDA_STANDARD = "17"
 CUPDLPX_BUILD_PYTHON = "ON"
 CUPDLPX_BUILD_STATIC_LIB = "ON"

src/cli.c

@@ -102,21 +102,6 @@ void save_solver_summary(const cupdlpx_result_t *result, const char *output_dir,
     }
     fprintf(outfile, "Termination Reason: %s\n",
             termination_reason_to_string(result->termination_reason));
-    fprintf(outfile, "Runtime (sec): %e\n", result->cumulative_time_sec);
-    fprintf(outfile, "Iterations Count: %d\n", result->total_count);
-    fprintf(outfile, "Primal Objective Value: %e\n",
-            result->primal_objective_value);
-    fprintf(outfile, "Dual Objective Value: %e\n", result->dual_objective_value);
-    fprintf(outfile, "Relative Primal Residual: %e\n",
-            result->relative_primal_residual);
-    fprintf(outfile, "Relative Dual Residual: %e\n",
-            result->relative_dual_residual);
-    fprintf(outfile, "Absolute Objective Gap: %e\n", result->objective_gap);
-    fprintf(outfile, "Relative Objective Gap: %e\n",
-            result->relative_objective_gap);
-    fprintf(outfile, "Rows: %d\n", result->num_constraints);
-    fprintf(outfile, "Columns: %d\n", result->num_variables);
-    fprintf(outfile, "Nonzeros: %d\n", result->num_nonzeros);
     if (result->presolve_time > 0.0)
     {
         fprintf(outfile, "Presolve Status: %s\n", get_presolve_status_str(result->presolve_status));
@@ -142,6 +127,22 @@ void save_solver_summary(const cupdlpx_result_t *result, const char *output_dir,
         //     fprintf(outfile, "Postsolve Time (sec): %e\n", result->presolve_stats.time_postsolve);
         // }
     }
+    fprintf(outfile, "Precondition time (sec): %e\n", result->rescaling_time_sec);
+    fprintf(outfile, "Runtime (sec): %e\n", result->cumulative_time_sec);
+    fprintf(outfile, "Iterations Count: %d\n", result->total_count);
+    fprintf(outfile, "Primal Objective Value: %e\n",
+            result->primal_objective_value);
+    fprintf(outfile, "Dual Objective Value: %e\n", result->dual_objective_value);
+    fprintf(outfile, "Relative Primal Residual: %e\n",
+            result->relative_primal_residual);
+    fprintf(outfile, "Relative Dual Residual: %e\n",
+            result->relative_dual_residual);
+    fprintf(outfile, "Absolute Objective Gap: %e\n", result->objective_gap);
+    fprintf(outfile, "Relative Objective Gap: %e\n",
+            result->relative_objective_gap);
+    fprintf(outfile, "Rows: %d\n", result->num_constraints);
+    fprintf(outfile, "Columns: %d\n", result->num_variables);
+    fprintf(outfile, "Nonzeros: %d\n", result->num_nonzeros);
     if (result->feasibility_polishing_time > 0.0)
     {
         fprintf(outfile, "Feasibility Polishing Time (sec): %e\n", result->feasibility_polishing_time);