fdurastante/psblas3
1
0
Fork 0
mirror of https://github.com/sfilippone/psblas3.git synced 2026-10-20 12:52:14 +00:00
Code Issues Packages Projects Releases Wiki Activity

Modified CSRG to work with latest versions; cusparse docs are unclear

Browse Source
This commit is contained in:
Salvatore Filippone
2023-12-15 03:41:19 -05:00
parent 6b65199afb
commit d28ea462d9
15 changed files with 368 additions and 92 deletions
Download Patch File Download Diff File Expand all files Collapse all files
cuda/impl/psb_c_cuda_csrg_to_gpu.F90
+61 -2
View File
@@ -227,7 +227,7 @@ subroutine psb_c_cuda_csrg_to_gpu(a,info,nzrm)
endif
#else
#elif 0
if (a%is_unit()) then
!
@@ -308,7 +308,66 @@ subroutine psb_c_cuda_csrg_to_gpu(a,info,nzrm)
!!$ if ((info == 0) .and. a%is_triangle()) then
!!$ info = CSRGDeviceCsrsmAnalysis(a%deviceMat)
!!$ end if
#else
if (a%is_unit()) then
!
! CUSPARSE has the habit of storing the diagonal and then ignoring,
! whereas we do not store it. Hence this adapter code.
!
nzdi = nz + m
if (info == 0) info = CSRGDeviceAlloc(a%deviceMat,m,n,nzdi)
if (info == 0) then
if (a%is_unit()) then
info = CSRGDeviceSetMatDiagType(a%deviceMat,cusparse_diag_type_unit)
else
info = CSRGDeviceSetMatDiagType(a%deviceMat,cusparse_diag_type_non_unit)
end if
end if
!!! We are explicitly adding the diagonal
if ((info == 0) .and. a%is_triangle()) then
if ((info == 0).and.a%is_upper()) then
info = CSRGDeviceSetMatFillMode(a%deviceMat,cusparse_fill_mode_upper)
else
info = CSRGDeviceSetMatFillMode(a%deviceMat,cusparse_fill_mode_lower)
end if
end if
if (info == 0) allocate(irpdi(m+1),jadi(nzdi),valdi(nzdi),stat=info)
if (info == 0) then
irpdi(1) = 1
if (a%is_triangle().and.a%is_upper()) then
do i=1,m
j = irpdi(i)
jadi(j) = i
valdi(j) = cone
nrz = a%irp(i+1)-a%irp(i)
jadi(j+1:j+nrz) = a%ja(a%irp(i):a%irp(i+1)-1)
valdi(j+1:j+nrz) = a%val(a%irp(i):a%irp(i+1)-1)
irpdi(i+1) = j + nrz + 1
! write(0,*) 'Row ',i,' : ',irpdi(i:i+1),':',jadi(j:j+nrz),valdi(j:j+nrz)
end do
else
do i=1,m
j = irpdi(i)
nrz = a%irp(i+1)-a%irp(i)
jadi(j+0:j+nrz-1) = a%ja(a%irp(i):a%irp(i+1)-1)
valdi(j+0:j+nrz-1) = a%val(a%irp(i):a%irp(i+1)-1)
jadi(j+nrz) = i
valdi(j+nrz) = cone
irpdi(i+1) = j + nrz + 1
! write(0,*) 'Row ',i,' : ',irpdi(i:i+1),':',jadi(j:j+nrz),valdi(j:j+nrz)
end do
end if
end if
if (info == 0) info = CSRGHost2Device(a%deviceMat,m,n,nzdi,irpdi,jadi,valdi)
else
if (info == 0) info = CSRGDeviceAlloc(a%deviceMat,m,n,nz)
if (info == 0) info = CSRGHost2Device(a%deviceMat,m,n,nz,a%irp,a%ja,a%val)
endif
#endif
call a%set_sync()
cuda/impl/psb_d_cuda_csrg_to_gpu.F90
+1 -20
View File
@@ -308,6 +308,7 @@ subroutine psb_d_cuda_csrg_to_gpu(a,info,nzrm)
!!$ if ((info == 0) .and. a%is_triangle()) then
!!$ info = CSRGDeviceCsrsmAnalysis(a%deviceMat)
!!$ end if
#else
if (a%is_unit()) then
@@ -325,9 +326,7 @@ subroutine psb_d_cuda_csrg_to_gpu(a,info,nzrm)
end if
end if
!!! We are explicitly adding the diagonal
!! info = CSRGDeviceSetMatDiagType(a%deviceMat,cusparse_diag_type_non_unit)
if ((info == 0) .and. a%is_triangle()) then
!!$ info = CSRGDeviceSetMatType(a%deviceMat,cusparse_matrix_type_triangular)
if ((info == 0).and.a%is_upper()) then
info = CSRGDeviceSetMatFillMode(a%deviceMat,cusparse_fill_mode_upper)
else
@@ -366,24 +365,6 @@ subroutine psb_d_cuda_csrg_to_gpu(a,info,nzrm)
else
if (info == 0) info = CSRGDeviceAlloc(a%deviceMat,m,n,nz)
!info = CSRGDeviceSetMatType(a%deviceMat,cusparse_matrix_type_general)
!!$ if (info == 0) info = CSRGDeviceSetMatIndexBase(a%deviceMat,cusparse_index_base_one)
!!$ if (a%is_triangle()) then
!!$ if (info == 0) then
!!$ if (a%is_unit()) then
!!$ info = CSRGDeviceSetMatDiagType(a%deviceMat,cusparse_diag_type_unit)
!!$ else
!!$ info = CSRGDeviceSetMatDiagType(a%deviceMat,cusparse_diag_type_non_unit)
!!$ end if
!!$ end if
!!$ if ((info == 0) )then
!!$ if ((info == 0).and.a%is_upper()) then
!!$ info = CSRGDeviceSetMatFillMode(a%deviceMat,cusparse_fill_mode_upper)
!!$ else
!!$ info = CSRGDeviceSetMatFillMode(a%deviceMat,cusparse_fill_mode_lower)
!!$ end if
!!$ end if
!!$ end if
if (info == 0) info = CSRGHost2Device(a%deviceMat,m,n,nz,a%irp,a%ja,a%val)
endif
cuda/impl/psb_s_cuda_csrg_to_gpu.F90
+61 -2
View File
@@ -227,7 +227,7 @@ subroutine psb_s_cuda_csrg_to_gpu(a,info,nzrm)
endif
#else
#elif 0
if (a%is_unit()) then
!
@@ -308,7 +308,66 @@ subroutine psb_s_cuda_csrg_to_gpu(a,info,nzrm)
!!$ if ((info == 0) .and. a%is_triangle()) then
!!$ info = CSRGDeviceCsrsmAnalysis(a%deviceMat)
!!$ end if
#else
if (a%is_unit()) then
!
! CUSPARSE has the habit of storing the diagonal and then ignoring,
! whereas we do not store it. Hence this adapter code.
!
nzdi = nz + m
if (info == 0) info = CSRGDeviceAlloc(a%deviceMat,m,n,nzdi)
if (info == 0) then
if (a%is_unit()) then
info = CSRGDeviceSetMatDiagType(a%deviceMat,cusparse_diag_type_unit)
else
info = CSRGDeviceSetMatDiagType(a%deviceMat,cusparse_diag_type_non_unit)
end if
end if
!!! We are explicitly adding the diagonal
if ((info == 0) .and. a%is_triangle()) then
if ((info == 0).and.a%is_upper()) then
info = CSRGDeviceSetMatFillMode(a%deviceMat,cusparse_fill_mode_upper)
else
info = CSRGDeviceSetMatFillMode(a%deviceMat,cusparse_fill_mode_lower)
end if
end if
if (info == 0) allocate(irpdi(m+1),jadi(nzdi),valdi(nzdi),stat=info)
if (info == 0) then
irpdi(1) = 1
if (a%is_triangle().and.a%is_upper()) then
do i=1,m
j = irpdi(i)
jadi(j) = i
valdi(j) = sone
nrz = a%irp(i+1)-a%irp(i)
jadi(j+1:j+nrz) = a%ja(a%irp(i):a%irp(i+1)-1)
valdi(j+1:j+nrz) = a%val(a%irp(i):a%irp(i+1)-1)
irpdi(i+1) = j + nrz + 1
! write(0,*) 'Row ',i,' : ',irpdi(i:i+1),':',jadi(j:j+nrz),valdi(j:j+nrz)
end do
else
do i=1,m
j = irpdi(i)
nrz = a%irp(i+1)-a%irp(i)
jadi(j+0:j+nrz-1) = a%ja(a%irp(i):a%irp(i+1)-1)
valdi(j+0:j+nrz-1) = a%val(a%irp(i):a%irp(i+1)-1)
jadi(j+nrz) = i
valdi(j+nrz) = sone
irpdi(i+1) = j + nrz + 1
! write(0,*) 'Row ',i,' : ',irpdi(i:i+1),':',jadi(j:j+nrz),valdi(j:j+nrz)
end do
end if
end if
if (info == 0) info = CSRGHost2Device(a%deviceMat,m,n,nzdi,irpdi,jadi,valdi)
else
if (info == 0) info = CSRGDeviceAlloc(a%deviceMat,m,n,nz)
if (info == 0) info = CSRGHost2Device(a%deviceMat,m,n,nz,a%irp,a%ja,a%val)
endif
#endif
call a%set_sync()
cuda/impl/psb_z_cuda_csrg_to_gpu.F90
+61 -2
View File
@@ -227,7 +227,7 @@ subroutine psb_z_cuda_csrg_to_gpu(a,info,nzrm)
endif
#else
#elif 0
if (a%is_unit()) then
!
@@ -308,7 +308,66 @@ subroutine psb_z_cuda_csrg_to_gpu(a,info,nzrm)
!!$ if ((info == 0) .and. a%is_triangle()) then
!!$ info = CSRGDeviceCsrsmAnalysis(a%deviceMat)
!!$ end if
#else
if (a%is_unit()) then
!
! CUSPARSE has the habit of storing the diagonal and then ignoring,
! whereas we do not store it. Hence this adapter code.
!
nzdi = nz + m
if (info == 0) info = CSRGDeviceAlloc(a%deviceMat,m,n,nzdi)
if (info == 0) then
if (a%is_unit()) then
info = CSRGDeviceSetMatDiagType(a%deviceMat,cusparse_diag_type_unit)
else
info = CSRGDeviceSetMatDiagType(a%deviceMat,cusparse_diag_type_non_unit)
end if
end if
!!! We are explicitly adding the diagonal
if ((info == 0) .and. a%is_triangle()) then
if ((info == 0).and.a%is_upper()) then
info = CSRGDeviceSetMatFillMode(a%deviceMat,cusparse_fill_mode_upper)
else
info = CSRGDeviceSetMatFillMode(a%deviceMat,cusparse_fill_mode_lower)
end if
end if
if (info == 0) allocate(irpdi(m+1),jadi(nzdi),valdi(nzdi),stat=info)
if (info == 0) then
irpdi(1) = 1
if (a%is_triangle().and.a%is_upper()) then
do i=1,m
j = irpdi(i)
jadi(j) = i
valdi(j) = zone
nrz = a%irp(i+1)-a%irp(i)
jadi(j+1:j+nrz) = a%ja(a%irp(i):a%irp(i+1)-1)
valdi(j+1:j+nrz) = a%val(a%irp(i):a%irp(i+1)-1)
irpdi(i+1) = j + nrz + 1
! write(0,*) 'Row ',i,' : ',irpdi(i:i+1),':',jadi(j:j+nrz),valdi(j:j+nrz)
end do
else
do i=1,m
j = irpdi(i)
nrz = a%irp(i+1)-a%irp(i)
jadi(j+0:j+nrz-1) = a%ja(a%irp(i):a%irp(i+1)-1)
valdi(j+0:j+nrz-1) = a%val(a%irp(i):a%irp(i+1)-1)
jadi(j+nrz) = i
valdi(j+nrz) = zone
irpdi(i+1) = j + nrz + 1
! write(0,*) 'Row ',i,' : ',irpdi(i:i+1),':',jadi(j:j+nrz),valdi(j:j+nrz)
end do
end if
end if
if (info == 0) info = CSRGHost2Device(a%deviceMat,m,n,nzdi,irpdi,jadi,valdi)
else
if (info == 0) info = CSRGDeviceAlloc(a%deviceMat,m,n,nz)
if (info == 0) info = CSRGHost2Device(a%deviceMat,m,n,nz,a%irp,a%ja,a%val)
endif
#endif
call a%set_sync()