Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Improve noncontiguous data transfers #13

Merged
merged 20 commits into from
Jan 22, 2024
Merged

Improve noncontiguous data transfers #13

Show file tree
Hide file tree
Changes from 10 commits
Commits
Show all changes
20 commits
Select commit Hold shift + click to select a range
aa504dd
speed up multidimensional non-contiguous transfers
lukasm91 Nov 14, 2023
8c5f5f9
dispatch to correct 2d variant
lukasm91 Nov 14, 2023
0e86739
make cpu only working
lukasm91 Nov 15, 2023
cab2dac
license
lukasm91 Nov 16, 2023
02ed2d8
fix compilation
lukasm91 Nov 17, 2023
19b073e
use contiguous copies for size=1
lukasm91 Nov 17, 2023
09cd076
remove prints
lukasm91 Nov 17, 2023
e327d07
stupid mistake in cpu code
lukasm91 Nov 17, 2023
0c79835
remove prints
lukasm91 Nov 20, 2023
68726f6
make tests smaller
lukasm91 Nov 20, 2023
6074669
Compile & run tests with Intel & nvhpc
pmarguinaud Jan 17, 2024
4faa538
Merge branch philippes merge
lukasm91 Jan 19, 2024
229a3a1
some minor fixes after merge
lukasm91 Jan 19, 2024
7a23cfa
remove cudamemcpydevicetohost argument and add better comments
lukasm91 Jan 19, 2024
2a127d7
Remove redundant specification of cuda compiler flags
awnawab Jan 19, 2024
35c8aee
Add pure device-pointer support to COPY_2D
awnawab Jan 19, 2024
4d1a3c8
FIELD_XX_COPY now works again if ACC is enabled but CUDA is disabled
awnawab Jan 19, 2024
3bb735f
Only perform CUDASUCCESS checks for DEBUG builds
awnawab Jan 19, 2024
b129adf
Restore CUDASUCCESS checks for optimized builds
awnawab Jan 22, 2024
ca16c19
DEBUG is not needed anymore
lukasm91 Jan 22, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions LICENSE
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -176,6 +176,7 @@
END OF TERMS AND CONDITIONS

Copyright 2018- ECMWF
Copyright 2023- NVIDIA

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
Expand Down
7 changes: 5 additions & 2 deletions cmake/field_api_compile_options.cmake
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,8 +10,11 @@
if(CMAKE_Fortran_COMPILER_ID MATCHES PGI|NVIDIA|NVHPC)
ecbuild_add_fortran_flags("-Mlarge_arrays")
ecbuild_add_fortran_flags("-gopt")

ecbuild_add_fortran_flags("-Minfo=accel,all,ccff" BUILD DEBUG)
if(HAVE_ACC)
list(APPEND CMAKE_REQUIRED_LINK_OPTIONS "-cuda")
ecbuild_add_fortran_flags("-cuda")
ecbuild_add_fortran_flags("-Minfo=accel,all,ccff" BUILD DEBUG)
endif()

# These are architecture/compiler/offload-library specific options
# that should really be coming from external input
Expand Down
148 changes: 130 additions & 18 deletions field_RANKSUFF_module.fypp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,6 @@
#! (C) Copyright 2022- ECMWF.
#! (C) Copyright 2022- Meteo-France.
#! (C) Copyright 2023- NVIDIA
#!
#! This software is licensed under the terms of the Apache Licence Version 2.0
#! which can be obtained at http://www.apache.org/licenses/LICENSE-2.0.
Expand Down Expand Up @@ -323,15 +324,28 @@ CONTAINS
INTEGER (KIND=JPIM), INTENT(IN) :: KDIR
INTEGER (KIND=JPIM), OPTIONAL, INTENT(IN) :: QUEUE
REAL :: START, FINISH
INTEGER :: NEXT_LAST_CONTIGUOUS_DIMENSION

SELF%LAST_CONTIGUOUS_DIMENSION = ${ftn}$_GET_LAST_CONTIGUOUS_DIMENSION (SELF%PTR)
NEXT_LAST_CONTIGUOUS_DIMENSION = ${ftn}$_GET_LAST_CONTIGUOUS_DIMENSION (SELF%PTR, SELF%LAST_CONTIGUOUS_DIMENSION+1)

CALL CPU_TIME(START)
SELECT CASE (SELF%LAST_CONTIGUOUS_DIMENSION)
#:for d in range (ft.rank + 1)
CASE (${d}$)
CALL COPY_DIM${d}$_CONTIGUOUS (SELF%PTR, SELF%DEVPTR, KDIR, QUEUE)
#:for d1 in range (ft.rank)
CASE (${d1}$)
SELECT CASE (NEXT_LAST_CONTIGUOUS_DIMENSION)
#:for d2 in range (d1+1, ft.rank+1)
CASE (${d2}$)
CALL COPY_2D_DIM${d1}$_${d2}$_CONTIGUOUS (SELF%PTR, SELF%DEVPTR, KDIR, QUEUE)
#:endfor
CASE DEFAULT
ERROR STOP
END SELECT
#:endfor
CASE (${ft.rank}$)
CALL COPY_DIM${ft.rank}$_CONTIGUOUS (SELF%PTR, SELF%DEVPTR, KDIR, QUEUE)
CASE DEFAULT
ERROR STOP
END SELECT
CALL CPU_TIME(FINISH)
IF (KDIR == NH2D) THEN
Expand All @@ -342,7 +356,86 @@ CONTAINS

CONTAINS

#:for d in range (0, ft.rank+1)
#:for d1 in range (0, ft.rank)
#:for d2 in range (d1+1, ft.rank+1)
SUBROUTINE COPY_2D_DIM${d1}$_${d2}$_CONTIGUOUS (HST, DEV, KDIR, QUEUE)
#ifdef _OPENACC
USE OPENACC
USE CUDAFOR
#endif
${ft.type}$, POINTER :: HST (${ft.shape}$), DEV (${ft.shape}$)
INTEGER (KIND=JPIM), INTENT (IN) :: KDIR
INTEGER (KIND=JPIM), OPTIONAL, INTENT (IN) :: QUEUE
#ifdef _OPENACC
INTEGER (KIND=JPIM) :: IHST_PITCH, IDEV_PITCH, IRET
INTEGER (KIND=JPIM) :: IWIDTH, IHEIGHT, ISHP(${ft.rank+1}$)
#:if d2 < ft.rank
INTEGER :: ${', '.join (list (map (lambda i: 'J' + str (i+1), range (d2, ft.rank))))}$
#:endif
INTEGER(KIND=CUDA_STREAM_KIND) :: STREAM

ISHP(1) = 1
ISHP(2:) = SHAPE(HST)
IWIDTH = PRODUCT(ISHP(1:${d1+1}$))
IHEIGHT = PRODUCT(ISHP(${d1+2}$:${d2+1}$))

#:set lb = lambda arr, i: f'LBOUND({arr}, {i+1})'
#:set lbnds = lambda arr, start, end: [lb(arr, i) for i in range(start, end)]
#:set this_slice = lambda arr: ', '.join(lbnds(arr, 0, ft.rank))
#:set next_slice = lambda arr: ', '.join(lbnds(arr, 0, d1) + [lb(arr, d1)+'+1'] + lbnds(arr, d1+1, ft.rank))
IHST_PITCH = (LOC (HST(${next_slice('HST')}$)) - LOC (HST (${this_slice('HST')}$)))/KIND(HST)
IDEV_PITCH = (LOC (DEV(${next_slice('DEV')}$)) - LOC (DEV (${this_slice('DEV')}$)))/KIND(DEV)

#:for e in range (ft.rank, d2, -1)
${' ' * (ft.rank - e)}$DO J${e}$ = LBOUND (HST, ${e}$), UBOUND (HST, ${e}$)
#:endfor
#:set indent = ' ' * (ft.rank - d2 - 1)
#:set ar = lambda arr: ', '.join(lbnds(arr, 0, d2) + [f'J{i+1}' for i in range(d2, ft.rank)])
${indent}$ IF (KDIR == NH2D) THEN
${indent}$ !$acc host_data use_device (DEV)
${indent}$ IF(PRESENT(QUEUE)) THEN
${indent}$ CALL ACC_SET_CUDA_STREAM(QUEUE, STREAM)
${indent}$ IRET = CUDAMEMCPY2DASYNC (DEV (${ar('DEV')}$), IDEV_PITCH, &
${indent}$ & HST (${ar('HST')}$), IHST_PITCH, &
${indent}$ & IWIDTH, IHEIGHT, CUDAMEMCPYHOSTTODEVICE, &
${indent}$ & STREAM)
${indent}$ ELSE
${indent}$ IRET = CUDAMEMCPY2D (DEV (${ar('DEV')}$), IDEV_PITCH, &
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Two questions :

  • CUDAMEMCPY2DASYNC seems to require a CUDAMEMCPYHOSTTODEVICE/CUDAMEMCPYDEVICETOHOST argument, while CUDAMEMCPY2D does not; is it normal ?
  • the direction of transfers triggered by CUDAMEMCPY2DASYNC/CUDAMEMCPY2D is influenced by the order of HST/DEV : H2D with DEV as first argument, D2H with HST as first argument. Correct ?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

  • You are right, I can do this better. I had to fill in CUDAMEMCPYXTOY because I want to specify the stream, but the better way of course is to do STREAM=STREAM
  • yes, all memcpy are memcpy(dst, [...], src, [...]). So for H2D, you have DEV first, and for D2H you have HST first. this is true for the normal C memcpy, but also for all cudaMemcpy variants.

Thanks for doing the merge with master! I merged this back with my branch and did this minor change. I also slightly change the test quoting explicitly which function is supposed to be called.

${indent}$ & HST (${ar('HST')}$), IHST_PITCH, &
${indent}$ & IWIDTH, IHEIGHT)
${indent}$ ENDIF
${indent}$ !$acc end host_data
${indent}$ ELSEIF (KDIR == ND2H) THEN
${indent}$ !$acc host_data use_device (DEV)
${indent}$ IF(PRESENT(QUEUE)) THEN
${indent}$ CALL ACC_SET_CUDA_STREAM(QUEUE, STREAM)
${indent}$ IRET = CUDAMEMCPY2DASYNC (HST (${ar('HST')}$), IHST_PITCH, &
${indent}$ & DEV (${ar('DEV')}$), IDEV_PITCH, &
${indent}$ & IWIDTH, IHEIGHT, CUDAMEMCPYDEVICETOHOST, &
${indent}$ & STREAM)
${indent}$ ELSE
${indent}$ IRET = CUDAMEMCPY2D (HST (${ar('HST')}$), IHST_PITCH, &
${indent}$ & DEV (${ar('DEV')}$), IDEV_PITCH, &
${indent}$ & IWIDTH, IHEIGHT)
${indent}$ ENDIF
${indent}$ !$acc end host_data
${indent}$ IF (IRET /= CUDASUCCESS) STOP 1
${indent}$ ENDIF
#:for e in range (d2, ft.rank)
${' ' * (ft.rank - e - 1)}$ENDDO
#:endfor
#else
IF (KDIR == NH2D) THEN
DEV = HST
ELSE
HST = DEV
ENDIF
#endif
END SUBROUTINE COPY_2D_DIM${d1}$_${d2}$_CONTIGUOUS
#:endfor
#:endfor

#:for d in range (ft.rank, ft.rank+1)
SUBROUTINE COPY_DIM${d}$_CONTIGUOUS (HST, DEV, KDIR, QUEUE)
#ifdef _OPENACC
USE OPENACC
Expand All @@ -351,13 +444,15 @@ CONTAINS
INTEGER (KIND=JPIM), INTENT (IN) :: KDIR
INTEGER (KIND=JPIM), OPTIONAL, INTENT (IN) :: QUEUE
INTEGER (KIND=JPIM) :: ISIZE
INTEGER :: ${', '.join (['J'] + list (map (lambda i: 'J' + str (i+1), range (d, ft.rank))))}$
#:if d != ft.rank
INTEGER :: ${', '.join (list (map (lambda i: 'J' + str (i+1), range (d, ft.rank))))}$
#:endif

#:for e in range (ft.rank, d, -1)
${' ' * (ft.rank - e)}$DO J${e}$ = LBOUND (HST, ${e}$), UBOUND (HST, ${e}$)
#:endfor
#:set ar = ', '.join ([':'] * d + list (map (lambda i: 'J' + str (i+1), range (d, ft.rank))))
#:set indent = ' ' * (ft.rank - e)
#:set indent = ' ' * (ft.rank - d)
#:if d == 0
${indent}$ ISIZE = KIND (HST)
#:else
Expand Down Expand Up @@ -605,18 +700,24 @@ CONTAINS

#:for ft in fieldTypeList
#:set ftn = ft.name
INTEGER (KIND=JPIM) FUNCTION ${ftn}$_GET_LAST_CONTIGUOUS_DIMENSION (PTR) RESULT (JDIM)
INTEGER (KIND=JPIM) FUNCTION ${ftn}$_GET_LAST_CONTIGUOUS_DIMENSION (PTR, AFTER) RESULT (JDIM)
${ft.type}$, POINTER :: PTR (${ft.shape}$)
INTEGER*8 :: ISTRIDE (${ft.rank}$)
INTEGER (KIND=JPIM) :: J, LB(${ft.rank}$)
INTEGER (KIND=JPIM), OPTIONAL :: AFTER
INTEGER*8 :: IPREVIOUS_STRIDE, ITHIS_STRIDE, ISIZE
INTEGER (KIND=JPIM) :: J, LB(${ft.rank}$), IAFTER

! assume that dimension all dimensions before IAFTER are contiguous...
IF (PRESENT(AFTER)) THEN
IAFTER = AFTER
ELSE
IAFTER = 0
ENDIF

LB = LBOUND(PTR)
ISTRIDE (1) = KIND (PTR)
DO J = 2, ${ft.rank}$
ISTRIDE (J) = ISTRIDE (J-1) * SIZE (PTR, J-1)
ENDDO
IF (IAFTER == 0) THEN
IPREVIOUS_STRIDE = KIND (PTR)
ENDIF

JDIM = 0
#:for d in range (1, ft.rank+1)
#:set ind0 = ""
#:set ind1 = ""
Expand All @@ -626,13 +727,24 @@ CONTAINS
#:endfor
#:set ind0 = ind0[:-2]
#:set ind1 = ind1[:-2]
IF (LOC (PTR (${ind1}$)) - LOC (PTR (${ind0}$)) /= ISTRIDE (${d}$)) THEN
RETURN
ITHIS_STRIDE = LOC (PTR (${ind1}$)) - LOC (PTR (${ind0}$))
IF (IAFTER < ${d}$) THEN
#:if d == 1
ISIZE = 1
#:else
ISIZE = SIZE(PTR, ${d-1}$)
#:endif
IF (SIZE(PTR, ${d}$) /= 1 .AND. IPREVIOUS_STRIDE * ISIZE /= ITHIS_STRIDE) THEN
JDIM = ${d-1}$
RETURN
ENDIF
IPREVIOUS_STRIDE = IPREVIOUS_STRIDE * ISIZE
ELSE IF (IAFTER == ${d}$) THEN
IPREVIOUS_STRIDE = ITHIS_STRIDE
ENDIF

JDIM = ${d}$

#:endfor
JDIM = ${ft.rank}$
END FUNCTION ${ftn}$_GET_LAST_CONTIGUOUS_DIMENSION
#:endfor

Expand Down
1 change: 1 addition & 0 deletions tests/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,6 +53,7 @@ list(APPEND TEST_FILES
init_wrapper_gpu.F90
init_wrapper_lbounds.F90
init_wrapper_non_contiguous.F90
init_wrapper_non_contiguous_multi.F90
no_transfer_get_device.F90
no_transfer_get_host.F90
pointer_to_owner_wrapper.F90
Expand Down
Loading