From 5375112eaf53e796d87c0abd6bd3282737f75725 Mon Sep 17 00:00:00 2001 From: shenw Date: Mon, 22 Jun 2026 21:00:21 +0800 Subject: [PATCH] Add conjugate gradient linear solver --- CMakeLists.txt | 13 +++ src/solver/fem_linear_solver_cg.f90 | 115 ++++++++++++++++++++++ tests/module/test_linear_solver_cg.f90 | 128 +++++++++++++++++++++++++ 3 files changed, 256 insertions(+) create mode 100644 src/solver/fem_linear_solver_cg.f90 create mode 100644 tests/module/test_linear_solver_cg.f90 diff --git a/CMakeLists.txt b/CMakeLists.txt index 23bcbf2..f53ddc0 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -47,6 +47,7 @@ add_library(femlite_core STATIC src/element/fem_element_mass.f90 src/element/fem_element_internal.f90 src/solver/fem_sparse_csr.f90 + src/solver/fem_linear_solver_cg.f90 src/coupling/fem_coupling.f90 src/boundary/fem_boundary.f90 src/output/fem_output_csv.f90 @@ -119,6 +120,13 @@ add_executable(test_sparse_csr target_link_libraries(test_sparse_csr PRIVATE femlite_core) femlite_apply_fortran_options(test_sparse_csr) +add_executable(test_linear_solver_cg + tests/module/test_linear_solver_cg.f90 +) + +target_link_libraries(test_linear_solver_cg PRIVATE femlite_core) +femlite_apply_fortran_options(test_linear_solver_cg) + add_executable(perf_p1 tests/performance/perf_p1.f90 ) @@ -160,3 +168,8 @@ add_test(NAME sparse_csr COMMAND test_sparse_csr) set_tests_properties(sparse_csr PROPERTIES WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}" ) + +add_test(NAME linear_solver_cg COMMAND test_linear_solver_cg) +set_tests_properties(linear_solver_cg PROPERTIES + WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}" +) diff --git a/src/solver/fem_linear_solver_cg.f90 b/src/solver/fem_linear_solver_cg.f90 new file mode 100644 index 0000000..e4f9bd6 --- /dev/null +++ b/src/solver/fem_linear_solver_cg.f90 @@ -0,0 +1,115 @@ +module fem_linear_solver_cg + use fem_precision + use fem_error + use fem_sparse_csr + implicit none + + private + + public :: linear_cg_solve + +contains + + subroutine linear_cg_solve(a, b, x, max_iter, tol, iter, res_norm, ierr) + type(csr_matrix), intent(in) :: a + real(rk), intent(in) :: b(:) + real(rk), intent(inout) :: x(:) + integer, intent(in) :: max_iter + real(rk), intent(in) :: tol + integer, intent(out) :: iter + real(rk), intent(out) :: res_norm + integer, intent(out) :: ierr + real(rk), allocatable :: r(:) + real(rk), allocatable :: p(:) + real(rk), allocatable :: ap(:) + real(rk) :: alpha + real(rk) :: beta + real(rk) :: denom + real(rk) :: norm_b + real(rk) :: rsold + real(rk) :: rsnew + real(rk) :: threshold + integer :: n + integer :: stat + + iter = 0 + res_norm = ZERO + + if (a%nrow .le. 0 .or. a%ncol .le. 0) then + call error_set(FEM_ERR_SOLVER, 'CG solver received an empty matrix', ierr) + return + end if + if (a%nrow .ne. a%ncol) then + call error_set(FEM_ERR_SOLVER, 'CG solver requires a square matrix', ierr) + return + end if + if (size(b) .ne. a%nrow .or. size(x) .ne. a%ncol) then + call error_set(FEM_ERR_SOLVER, 'CG solver vector size mismatch', ierr) + return + end if + if (max_iter .le. 0) then + call error_set(FEM_ERR_SOLVER, 'CG solver max_iter must be positive', ierr) + return + end if + if (tol .le. ZERO) then + call error_set(FEM_ERR_SOLVER, 'CG solver tolerance must be positive', ierr) + return + end if + + n = a%nrow + allocate(r(n), p(n), ap(n), stat=stat) + if (stat .ne. 0) then + call error_set(FEM_ERR_SOLVER, 'failed to allocate CG workspace', ierr) + return + end if + + call csr_matvec(a, x, ap, ierr) + if (ierr .ne. FEM_SUCCESS) go to 900 + + r = b - ap + p = r + norm_b = sqrt(dot_product(b, b)) + threshold = tol * max(ONE, norm_b) + rsold = dot_product(r, r) + res_norm = sqrt(rsold) + if (res_norm .le. threshold) then + ierr = FEM_SUCCESS + go to 900 + end if + + do iter = 1, max_iter + call csr_matvec(a, p, ap, ierr) + if (ierr .ne. FEM_SUCCESS) go to 900 + + denom = dot_product(p, ap) + if (denom .le. tiny(ONE)) then + call error_set(FEM_ERR_SOLVER, 'CG solver detected non-positive curvature', ierr) + go to 900 + end if + + alpha = rsold / denom + x = x + alpha * p + r = r - alpha * ap + + rsnew = dot_product(r, r) + res_norm = sqrt(rsnew) + if (res_norm .le. threshold) then + ierr = FEM_SUCCESS + go to 900 + end if + + beta = rsnew / rsold + p = r + beta * p + rsold = rsnew + end do + + iter = max_iter + call error_set(FEM_ERR_SOLVER, 'CG solver did not converge', ierr) + +900 continue + if (allocated(r)) deallocate(r) + if (allocated(p)) deallocate(p) + if (allocated(ap)) deallocate(ap) + end subroutine linear_cg_solve + +end module fem_linear_solver_cg diff --git a/tests/module/test_linear_solver_cg.f90 b/tests/module/test_linear_solver_cg.f90 new file mode 100644 index 0000000..11ab20e --- /dev/null +++ b/tests/module/test_linear_solver_cg.f90 @@ -0,0 +1,128 @@ +program test_linear_solver_cg + use fem_precision + use fem_error + use fem_sparse_csr + use fem_linear_solver_cg + implicit none + + real(rk), parameter :: TOL = 1.0e-9_rk + + call check_spd_solve() + call check_zero_rhs() + call check_nonconvergence() + +contains + + subroutine check_spd_solve() + type(csr_matrix) :: a + integer :: ierr + integer :: iter + real(rk) :: b(3) + real(rk) :: x(3) + real(rk) :: res_norm + + call make_tridiagonal(a, ierr) + if (ierr .ne. FEM_SUCCESS) call fail(1) + + b = (/ ZERO, ZERO, 4.0_rk /) + x = ZERO + call linear_cg_solve(a, b, x, 20, 1.0e-12_rk, iter, res_norm, ierr) + if (ierr .ne. FEM_SUCCESS) call fail(2) + if (iter .le. 0 .or. iter .gt. 3) call fail(3) + if (res_norm .gt. 1.0e-10_rk) call fail(4) + call check_close(x(1), ONE, 5) + call check_close(x(2), 2.0_rk, 6) + call check_close(x(3), 3.0_rk, 7) + + call csr_free(a, ierr) + if (ierr .ne. FEM_SUCCESS) call fail(8) + end subroutine check_spd_solve + + subroutine check_zero_rhs() + type(csr_matrix) :: a + integer :: ierr + integer :: iter + real(rk) :: b(3) + real(rk) :: x(3) + real(rk) :: res_norm + + call make_tridiagonal(a, ierr) + if (ierr .ne. FEM_SUCCESS) call fail(20) + + b = ZERO + x = ZERO + call linear_cg_solve(a, b, x, 20, 1.0e-12_rk, iter, res_norm, ierr) + if (ierr .ne. FEM_SUCCESS) call fail(21) + if (iter .ne. 0) call fail(22) + if (res_norm .gt. ZERO) call fail(23) + if (maxval(abs(x)) .gt. ZERO) call fail(24) + + call csr_free(a, ierr) + if (ierr .ne. FEM_SUCCESS) call fail(25) + end subroutine check_zero_rhs + + subroutine check_nonconvergence() + type(csr_matrix) :: a + integer :: ierr + integer :: iter + real(rk) :: b(3) + real(rk) :: x(3) + real(rk) :: res_norm + + call make_tridiagonal(a, ierr) + if (ierr .ne. FEM_SUCCESS) call fail(30) + + b = (/ ZERO, ZERO, 4.0_rk /) + x = ZERO + call linear_cg_solve(a, b, x, 1, 1.0e-12_rk, iter, res_norm, ierr) + if (ierr .eq. FEM_SUCCESS) call fail(31) + if (iter .ne. 1) call fail(32) + if (res_norm .le. ZERO) call fail(33) + + call csr_free(a, ierr) + if (ierr .ne. FEM_SUCCESS) call fail(34) + end subroutine check_nonconvergence + + subroutine make_tridiagonal(a, ierr) + type(csr_matrix), intent(inout) :: a + integer, intent(out) :: ierr + integer :: row_ptr(4) + integer :: col_ind(7) + + call csr_init(a) + row_ptr = (/ 1, 3, 6, 8 /) + col_ind = (/ 1, 2, 1, 2, 3, 2, 3 /) + + call csr_create_pattern(a, 3, 3, row_ptr, col_ind, ierr) + if (ierr .ne. FEM_SUCCESS) return + call csr_add_value(a, 1, 1, 2.0_rk, ierr) + if (ierr .ne. FEM_SUCCESS) return + call csr_add_value(a, 1, 2, -ONE, ierr) + if (ierr .ne. FEM_SUCCESS) return + call csr_add_value(a, 2, 1, -ONE, ierr) + if (ierr .ne. FEM_SUCCESS) return + call csr_add_value(a, 2, 2, 2.0_rk, ierr) + if (ierr .ne. FEM_SUCCESS) return + call csr_add_value(a, 2, 3, -ONE, ierr) + if (ierr .ne. FEM_SUCCESS) return + call csr_add_value(a, 3, 2, -ONE, ierr) + if (ierr .ne. FEM_SUCCESS) return + call csr_add_value(a, 3, 3, 2.0_rk, ierr) + end subroutine make_tridiagonal + + subroutine check_close(actual, expected, code) + real(rk), intent(in) :: actual + real(rk), intent(in) :: expected + integer, intent(in) :: code + + if (abs(actual - expected) .gt. TOL) call fail(code) + end subroutine check_close + + subroutine fail(code) + integer, intent(in) :: code + + write(*,*) 'test_linear_solver_cg failed:', code + stop 1 + end subroutine fail + +end program test_linear_solver_cg