78fbedc9a1
- Added benchmark_transpose.f90 to measure blocked transpose performance. - Added tune_blocksize.py to auto-detect grid sizes, benchmark valid block size candidates, and apply the optimal value. - Updated makefile to support benchmark targets and default BLOCKSIZE=16 based on autotuning. - Saved tuning results and system specifications to tuning_results.md.
73 lines
1.5 KiB
Fortran
73 lines
1.5 KiB
Fortran
program benchmark_transpose
|
|
use m_parameters
|
|
use m_calculate
|
|
implicit none
|
|
|
|
real*8, allocatable, dimension(:,:,:) :: aaa, bbb
|
|
integer :: ierr, iter, num_iters
|
|
real*8 :: t1, t2
|
|
character(len=32) :: arg
|
|
|
|
! Read command-line arguments for grid sizes
|
|
if (command_argument_count() >= 3) then
|
|
call get_command_argument(1, arg)
|
|
read(arg, *) nxp
|
|
call get_command_argument(2, arg)
|
|
read(arg, *) nyp
|
|
call get_command_argument(3, arg)
|
|
read(arg, *) nzp
|
|
else
|
|
! Default matching test_calculate
|
|
nxp = 512
|
|
nyp = 256
|
|
nzp = 256
|
|
end if
|
|
|
|
! Read iteration count
|
|
num_iters = 10
|
|
if (command_argument_count() >= 4) then
|
|
call get_command_argument(4, arg)
|
|
read(arg, *) num_iters
|
|
end if
|
|
|
|
l_0 = 2.0
|
|
hyp = l_0 * pi / REAL(nyp)
|
|
hxp = hyp
|
|
hzp = hyp
|
|
|
|
allocate(aaa(nxp, nyp, nzp), stat=ierr)
|
|
if (ierr /= 0) then
|
|
print *, "Error allocating array aaa"
|
|
stop 1
|
|
end if
|
|
|
|
allocate(bbb(nxp, nyp, nzp), stat=ierr)
|
|
if (ierr /= 0) then
|
|
print *, "Error allocating array bbb"
|
|
stop 1
|
|
end if
|
|
|
|
! Initialize dummy data
|
|
call random_number(aaa)
|
|
|
|
call m_calculate_init
|
|
|
|
! Warmup run
|
|
call ddx(bbb, aaa)
|
|
|
|
! Benchmark runs
|
|
call cpu_time(t1)
|
|
do iter = 1, num_iters
|
|
call ddx(bbb, aaa)
|
|
end do
|
|
call cpu_time(t2)
|
|
|
|
! Print execution time per call in seconds
|
|
print '(F10.6)', (t2 - t1) / num_iters
|
|
|
|
call m_calculate_finalize
|
|
|
|
deallocate(aaa)
|
|
deallocate(bbb)
|
|
|
|
end program benchmark_transpose
|