Will the next release of “pgf90” ( 11.0 ) support, acc region on derive type on Fermi GPU ?
PROGRAM test_derive_type
IMPLICIT NONE
INTEGER , PARAMETER :: N=64
TYPE GRID
REAL , DIMENSION(N) :: H2O,O3
END TYPE GRID
TYPE(grid) :: tgrid
!$acc region
tgrid%H2O(:) = 1.0
tgrid%O3(:) = 1.0 / ( 1.0 * N )
!$acc end region
print *, "TGRID=", tgrid%H2O(N) , tgrid%O3(N)
END PROGRAM test_derive_type
With the pgf90-10.9 :
pgf90 -V
pgf90 10.9-0 64-bit target on x86-64 Linux -tp nehalem-64
pgf90 -ta=nvidia,cuda3.1,cc20 -Minfo=accel simple_acc_derive_type.f90 -o simple_acc_derive_type
test_derive_type:
13, Accelerator region ignored
14, Accelerator restriction: struct/union/derived type variables not yet supported: tgrid%h2o
Accelerator restriction: struct/union/derived type variables not yet supported: tgrid%o3
Accelerator restriction: struct/member references are not yet supported
The problem here is not that derived types aren’t supported (they are) rather accelerating an array within a single derived type is not. Changing the code to make an array of derived types will allow the code to work.
% cat dt.f90
PROGRAM test_derive_type
IMPLICIT NONE
INTEGER , PARAMETER :: N=64
INTEGER :: i
TYPE GRID
REAL :: H2O,O3
END TYPE GRID
TYPE(grid),dimension(N) :: tgrid
!$acc region
tgrid(:)%H2O = 1.0
tgrid(:)%O3 = 1.0 / ( 1.0 * N )
!$acc end region
print *, "TGRID=", tgrid(N)%H2O , tgrid(N)%O3
END PROGRAM test_derive_type
% pgf90 -ta=nvidia -Minfo=accel dt.f90 -V10.9
test_derive_type:
15, Generating copyout(tgrid(1:64))
Generating compute capability 1.0 binary
Generating compute capability 1.3 binary
16, Loop is parallelizable
Accelerator kernel generated
16, !$acc do parallel, vector(64)
CC 1.0 : 3 registers; 20 shared, 16 constant, 0 local memory bytes; 66 occupancy
CC 1.3 : 3 registers; 20 shared, 16 constant, 0 local memory bytes; 50 occupancy
% a.out
TGRID= 1.000000 1.5625000E-02
The H2O data are not stride1 , and that will affect the performance because in our code and the “true” derived type contain about 70 chemicals species , and 100th of reactions are computed in the subroutines .
It’s why we use the derived type, with array as member’s , to pass only 1 argument to subroutine and not 100 of them …
I will try your suggestion and report the performance of the 2 differents forms .
This ‘basic’ functionality is not already supported with the PGI 10.8 version .
With cuda4.0 a lot a new functionally have been added for the management of
pointer & memory between CPU & GPU . I don’t see were is the ‘difficult point’ with this problem ?
So my question is when it would be ?
Do I have to post this question in an other forum <=> “Wish List” ??
tee error_cuda
PGF90/x86-64 Linux 11.8-0
Copyright 1989-2000, The Portland Group, Inc. All Rights Reserved.
Copyright 2000-2011, STMicroelectronics, Inc. All Rights Reserved.
PGF90/x86-64 Linux 11.8-0
Copyright 1989-2000, The Portland Group, Inc. All Rights Reserved.
Copyright 2000-2011, STMicroelectronics, Inc. All Rights Reserved.
PGF90-W-0155-Accelerator region ignored; see -Minfo messages (simple_acc_derive_type.f90: 13)
test_derive_type:
13, Intensity = 0.0
Accelerator region ignored
14, Intensity = 0.0
Accelerator restriction: struct/union/derived type variables not yet supported: tgrid%h2o
Accelerator restriction: struct/union/derived type variables not yet supported: tgrid%o3
Accelerator restriction: struct/member references are not yet supported
I have read the PGI article on the Fortran 2003 Object-Oriented Programming …
… but the ACC part alway miss the simple “derive type” syntax for intensive computation .
For use it’s a priority as all our code <=> 1 millions lines of Fortran90 , use already “object-oriented” style with “stream computing” like “array syntax” .
In France & Europe we have now a new big Petaflopic BULLX cluster ‘CURIE’ with last Fermi GPU & PGI compiler <=> see here http://www-hpc.cea.fr/fr/complexe/tgcc-curie.htm
… but we could not use it in the ‘Grand Challenge Project’ because of this lack ‘basic’ functional .