进行LU分解时代入初始化矩阵数据存在精度丢失

David.C.J · 发表于 2024-6-16 20:40:19

[Fortran] 纯文本查看 复制代码

    program main
    integer,parameter::sp = selected_real_kind(15)
    real(kind=sp)::mat_1(3)
    real(kind=sp) matrix_A1(2,2)
    real(kind=sp) matrix_A(7,7),L(7,7),U(7,7)
    integer*4 n,i
    !call ieee_set_rounding_mode(ieee_nearest)       !设置浮点数的精度
    n = 7
    mat_1 = (/1.0,2.0,3.0/)
    matrix_A1 = reshape((/2.0,2.0,1.0,1.0/),(/2,2/))
    matrix_A(:,:) = 0.0d0
    matrix_A(1,1:7) = [1.30000000,0.09523810,0.38461538,3.00000000,6.00000000,10.00000000,6.00000000]

    !matrix_A(1,1) = 1.30000000_sp  !即使定义real*16后矩阵初始化的值也是存在精度丢失

    write(*,*)"sp精度位数",sp
    !write(*,"(7F13.8)") matrix_A(1,1:7)
    write(*,*)"----------输出A矩阵----------"
    write(*,"(7F13.8)") (matrix_A(i,:),i=1,7)

    pause   
    end

无论怎么修改selected_real_kind里面要求的精度，但是初始化matrix_A(1,1)得到的值都是1.29999995，而不是预期的1.30000000
单独初始化为matrix_A(1,1) = 1.30000000_sp 输出的值才能是对的
那么如何保证矩阵整个赋值初始化事后不丢失精度呢
还望大佬们指教

li913 · 发表于 2024-6-17 08:44:04

数字也有精度区别， matrix_A(1,1:7) = [1.30000000,0.09523810,0.38461538,3.00000000,6.00000000,10.00000000,6.00000000] 这里等号后面都是单精度，数字末尾加 _sp . 例如 1.30000000_sp,0.09523810_sp
详细可参考视频 https://www.bilibili.com/video/BV1XD4y1S7jz?p=3