如何解决为什么某些数组在 Fortran 中没有正确对齐以进行矢量化?
我正在尝试通过向量化 for 循环来改进 Fortran 77 代码。我对矢量化还很陌生,虽然我可以让代码进行矢量化,但优化报告告诉我,我的某些数组具有未对齐的访问权限。据我所知,这使得矢量化效率降低。 我已手动向数组添加填充以对齐数据,这似乎适用于我的大多数数组,但不适用于所有数组(请参阅示例代码)。
SUbroUTINE SOURCE
ParaMETER(NLInes=8)
ParaMETER(NLM=NLInes-1)
ParaMETER(NZ=18)
COMMON/TEST/h(-nlines:nlines+15,-nlines:nlines)
&,hr(-nlines:nlines+15,-nlines:nlines)
COMMON/PRO/P2(-nlm:nlm+1,-nlm:nlm)
dimension SNT(-nlm:nlm+1,-nlm:nlm,0:NZ+1)
DO iy=-nlm,nlm
DO ix=-nlm,nlm
H(ix,iy)=SNT(ix,iy,1)
& +SNT(ix,0)
& +SNT(ix,2)
HR(ix,iy)=P2(ix,iy)*H(ix,iy)
enddo
enddo
END
以及优化报告的相关部分:
LOOP BEGIN at SOURCE.FPP(13,8)
remark #15389: vectorization support: reference h(ix,iy) has unaligned access [ SOURCE.FPP(14,9) ]
remark #15388: vectorization support: reference snt(ix,1) has aligned access [ SOURCE.FPP(14,0) has aligned access [ SOURCE.FPP(14,2) has aligned access [ SOURCE.FPP(15,9) ]
remark #15389: vectorization support: reference hr(ix,iy) has unaligned access [ SOURCE.FPP(18,9) ]
remark #15388: vectorization support: reference p2(ix,iy) has aligned access [ SOURCE.FPP(18,9) ]
remark #15389: vectorization support: reference h(ix,9) ]
remark #15381: vectorization support: unaligned access used inside loop body
remark #15305: vectorization support: vector length 4
remark #15399: vectorization support: unroll factor set to 3
remark #15300: LOOP WAS VECTORIZED
remark #15448: unmasked aligned unit stride loads: 4
remark #15450: unmasked unaligned unit stride loads: 1
remark #15451: unmasked unaligned unit stride stores: 2
remark #15475: --- begin vector cost summary ---
remark #15476: scalar cost: 15
remark #15477: vector cost: 4.250
remark #15478: estimated potential speedup: 2.340
remark #15488: --- end vector cost summary ---
remark #25456: Number of Array Refs Scalar Replaced In Loop: 3
remark #25015: Estimate of max trip count of loop=1
LOOP END
为什么 h(ix,iy) 和 hr(ix,iy) 没有对齐访问?以及如何对齐它们以获得更好的矢量化?
可能还有相关信息:
Intel(R) Fortran Intel(R) 64 Compiler for applications running on Intel(R) 64,Version 19.0.0.117 Build 20180804
Compiler options: -qopt-report=5 -c -DLINUX -O3 -w -assume nounderscore -xcore-AVX2 -o SOURCE.o193
提前感谢您的时间和帮助!
版权声明:本文内容由互联网用户自发贡献,该文观点与技术仅代表作者本人。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如发现本站有涉嫌侵权/违法违规的内容, 请发送邮件至 dio@foxmail.com 举报,一经查实,本站将立刻删除。