Add GPU implementation of preconditioner - v2

CMakeLists.txt

@@ -8,7 +8,7 @@ project(cupdlpx LANGUAGES C CXX CUDA)
 
 set(CUPDLPX_VERSION_MAJOR 0)
 set(CUPDLPX_VERSION_MINOR 1)
-set(CUPDLPX_VERSION_PATCH 4)
+set(CUPDLPX_VERSION_PATCH 5)
 
 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 = "0.1.3"
+version = "0.1.5"
 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 = "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

@@ -123,6 +123,7 @@ void save_solver_summary(const cupdlpx_result_t *result, const char *output_dir,
     }
     fprintf(outfile, "Termination Reason: %s\n",
             termination_reason_tToString(result->termination_reason));
+    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",