https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95397
Kirill Chilikin <chilikin.k at gmail dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |chilikin.k at gmail dot com
--- Comment #4 from Kirill Chilikin <chilikin.k at gmail dot com> ---
I found another case of incorrect results depending on the presence of the "acc
loop vector" directive (and the original example also fails):
$ cat test1.f90
PROGRAM TEST
USE, INTRINSIC :: ISO_FORTRAN_ENV
IMPLICIT NONE
INTEGER, PARAMETER :: N1 = 32
INTEGER, PARAMETER :: N2 = 32
REAL(REAL64), DIMENSION(N1, N2) :: A
INTEGER I1, I2
!$ACC PARALLEL COPYOUT(A)
!$ACC LOOP WORKER
DO I2 = 1, N2
BLOCK
REAL(REAL64), DIMENSION(N1) :: V
!$ACC LOOP VECTOR
DO I1 = 1, N1
V(I1) = REAL(I1)
ENDDO
!$ACC LOOP VECTOR
DO I1 = 1, N1
A(I1,I2) = V(I1)
ENDDO
END BLOCK
ENDDO
!$ACC END PARALLEL
PRINT *, A(:, 1)
END PROGRAM
$ gfortran -o test1 test1.f90 -fopenacc -foffload=nvptx-none
-foffload-options="-misa=sm_35"
chilikin@comp1:/mnt/raid/chilikin/binp/belle2/analysis_physics/psipik/externals/test_offloading/vectorization$
./test1
1.0000000000000000 0.0000000000000000 0.0000000000000000
0.0000000000000000 0.0000000000000000 0.0000000000000000
0.0000000000000000 0.0000000000000000 0.0000000000000000
0.0000000000000000 0.0000000000000000 0.0000000000000000
0.0000000000000000 0.0000000000000000 0.0000000000000000
0.0000000000000000 0.0000000000000000 0.0000000000000000
0.0000000000000000 0.0000000000000000 0.0000000000000000
0.0000000000000000 0.0000000000000000 0.0000000000000000
0.0000000000000000 0.0000000000000000 0.0000000000000000
0.0000000000000000 0.0000000000000000 0.0000000000000000
0.0000000000000000 0.0000000000000000
If the first "acc loop vector" directive is removed (test2.f90), then
$ gfortran -o test2 test2.f90 -fopenacc -foffload=nvptx-none
-foffload-options="-misa=sm_35"
$ ./test2
1.0000000000000000 2.0000000000000000 3.0000000000000000
4.0000000000000000 5.0000000000000000 6.0000000000000000
7.0000000000000000 8.0000000000000000 9.0000000000000000
10.000000000000000 11.000000000000000 12.000000000000000
13.000000000000000 14.000000000000000 15.000000000000000
16.000000000000000 17.000000000000000 18.000000000000000
19.000000000000000 20.000000000000000 21.000000000000000
22.000000000000000 23.000000000000000 24.000000000000000
25.000000000000000 26.000000000000000 27.000000000000000
28.000000000000000 29.000000000000000 30.000000000000000
31.000000000000000 32.000000000000000
GPU is Nvidia GeForce GT 710, compiler is
$ gfortran -v
Using built-in specs.
COLLECT_GCC=gfortran
COLLECT_LTO_WRAPPER=/path/offloading/libexec/gcc/x86_64-pc-linux-gnu/14.2.0/lto-wrapper
OFFLOAD_TARGET_NAMES=nvptx-none
Target: x86_64-pc-linux-gnu
Configured with: /source_path/gcc-14.2.0/configure --prefix=/path/offloading
--with-gmp=/path --with-mpfr=/path --with-mpc=/path --with-isl=/path
--disable-multilib --enable-offload-targets=nvptx-none
Thread model: posix
Supported LTO compression algorithms: zlib zstd
gcc version 14.2.0 (GCC)