From 3586a2dbe3d805f3170e47e21ad675760219a13c Mon Sep 17 00:00:00 2001 From: Edoardo Aprà Date: Tue, 13 Jun 2023 13:43:39 -0700 Subject: NWChem: fix fftw3 variant + patch (#38250) --- .../builtin/packages/nwchem/fftw_splans.patch | 643 +++++++++++++++++++++ var/spack/repos/builtin/packages/nwchem/package.py | 11 + 2 files changed, 654 insertions(+) create mode 100644 var/spack/repos/builtin/packages/nwchem/fftw_splans.patch (limited to 'var') diff --git a/var/spack/repos/builtin/packages/nwchem/fftw_splans.patch b/var/spack/repos/builtin/packages/nwchem/fftw_splans.patch new file mode 100644 index 0000000000..3eed13b855 --- /dev/null +++ b/var/spack/repos/builtin/packages/nwchem/fftw_splans.patch @@ -0,0 +1,643 @@ +From b4ec4ade1af434bc80470d6874aebf6fdcd12489 Mon Sep 17 00:00:00 2001 +From: Eric Bylaska +Date: Thu, 8 Jun 2023 14:21:43 -0700 +Subject: [PATCH] changing old fortran interface for fftw to be sfftw for + single precision...EJB + +--- + src/nwpw/nwpwlib/D3dB/D3dB-new.F | 3 +- + src/nwpw/nwpwlib/D3dB/D3dBs-new.F | 69 +++++++++++++++--------------- + src/nwpw/nwpwlib/D3dB/D3dBs_pfft.F | 64 +++++++++++++-------------- + src/nwpw/nwpwlib/D3dB/GNUmakefile | 4 +- + 4 files changed, 72 insertions(+), 68 deletions(-) + +diff --git a/src/nwpw/nwpwlib/D3dB/D3dB-new.F b/src/nwpw/nwpwlib/D3dB/D3dB-new.F +index fb00838d2c..2159707831 100644 +--- a/src/nwpw/nwpwlib/D3dB/D3dB-new.F ++++ b/src/nwpw/nwpwlib/D3dB/D3dB-new.F +@@ -1183,7 +1183,8 @@ subroutine D3dB_fft_init(nb) + common / D3dB_fft / tmpx,tmpy,tmpz + + #ifdef FFTW3 +-#include "fftw3.fh" ++!#include "fftw3.fh" ++#include "fftw3.f" + integer Atest(2),nxh,nxhz,nxhy + #endif + +diff --git a/src/nwpw/nwpwlib/D3dB/D3dBs-new.F b/src/nwpw/nwpwlib/D3dB/D3dBs-new.F +index 90127b7ff1..8ddfddad34 100644 +--- a/src/nwpw/nwpwlib/D3dB/D3dBs-new.F ++++ b/src/nwpw/nwpwlib/D3dB/D3dBs-new.F +@@ -195,7 +195,8 @@ subroutine D3dBs_fft_init(nb) + common / D3dBs_fft / tmpx,tmpy,tmpz + + #ifdef FFTW3 +-#include "fftw3.fh" ++!#include "fftw3.fh" ++#include "fftw3.f" + integer Atest(2),nxh,nxhz,nxhy + #endif + +@@ -248,48 +249,48 @@ subroutine D3dBs_fft_init(nb) + > 'Atest',Atest(2),Atest(1))) + > call errquit('D3dBs_fft_init:out of heap memory',0,MA_ERR) + +- call fftw_plan_many_dft(splans(1,nb),1,nz(nb),nxh, ++ call sfftw_plan_many_dft(splans(1,nb),1,nz(nb),nxh, + > scpl_mb(Atest(1)),nxhz,nxh,1, + > scpl_mb(Atest(1)),nxhz,nxh,1, + > ibackward,FFTW_EXHAUSTIVE) +- call fftw_plan_many_dft(splans(2,nb),1,ny(nb),nxh, ++ call sfftw_plan_many_dft(splans(2,nb),1,ny(nb),nxh, + > scpl_mb(Atest(1)),nxhy,nxh,1, + > scpl_mb(Atest(1)),nxhy,nxh,1, + > ibackward,FFTW_EXHAUSTIVE) +- call fftw_plan_many_dft_c2r(splans(3,nb),1,nx(nb), ++ call sfftw_plan_many_dft_c2r(splans(3,nb),1,nx(nb), + > ny(nb)*nq(nb), + > scpl_mb(Atest(1)),nxh *ny(nb)*nq(nb),1,nxh, + > scpl_mb(Atest(1)),(nx(nb)+2)*ny(nb)*nq(nb),1,nx(nb)+2, + > FFTW_EXHAUSTIVE) + +- call fftw_plan_many_dft_r2c(splans(4,nb),1,nx(nb), ++ call sfftw_plan_many_dft_r2c(splans(4,nb),1,nx(nb), + > ny(nb)*nq(nb), + > scpl_mb(Atest(1)),(nx(nb)+2)*ny(nb)*nq(nb),1,nx(nb)+2, + > scpl_mb(Atest(1)),nxh *ny(nb)*nq(nb),1,nxh, + > FFTW_EXHAUSTIVE) +- call fftw_plan_many_dft(splans(5,nb),1,ny(nb),nxh, ++ call sfftw_plan_many_dft(splans(5,nb),1,ny(nb),nxh, + > scpl_mb(Atest(1)),nxhy,nxh,1, + > scpl_mb(Atest(1)),nxhy,nxh,1, + > iforward,FFTW_EXHAUSTIVE) +- call fftw_plan_many_dft(splans(6,nb),1,nz(nb),nxh, ++ call sfftw_plan_many_dft(splans(6,nb),1,nz(nb),nxh, + > scpl_mb(Atest(1)),nxhz,nxh,1, + > scpl_mb(Atest(1)),nxhz,nxh,1, + > iforward,FFTW_EXHAUSTIVE) + +- call fftw_plan_many_dft(splans(7,nb),1,nz(nb),1, ++ call sfftw_plan_many_dft(splans(7,nb),1,nz(nb),1, + > scpl_mb(Atest(1)),nxhz,nxh,1, + > scpl_mb(Atest(1)),nxhz,nxh,1, + > ibackward,FFTW_EXHAUSTIVE) +- call fftw_plan_many_dft(splans(8,nb),1,ny(nb),1, ++ call sfftw_plan_many_dft(splans(8,nb),1,ny(nb),1, + > scpl_mb(Atest(1)),nxhy,nxh,1, + > scpl_mb(Atest(1)),nxhy,nxh,1, + > ibackward,FFTW_EXHAUSTIVE) + +- call fftw_plan_many_dft(splans(9,nb),1,ny(nb),1, ++ call sfftw_plan_many_dft(splans(9,nb),1,ny(nb),1, + > scpl_mb(Atest(1)),nxhy,nxh,1, + > scpl_mb(Atest(1)),nxhy,nxh,1, + > iforward,FFTW_EXHAUSTIVE) +- call fftw_plan_many_dft(splans(10,nb),1,nz(nb),1, ++ call sfftw_plan_many_dft(splans(10,nb),1,nz(nb),1, + > scpl_mb(Atest(1)),nxhz,nxh,1, + > scpl_mb(Atest(1)),nxhz,nxh,1, + > iforward,FFTW_EXHAUSTIVE) +@@ -305,48 +306,48 @@ subroutine D3dBs_fft_init(nb) + > Atest(2),Atest(1))) + > call errquit('D3dBs_fft_init:out of heap memory',0,MA_ERR) + +- call fftw_plan_many_dft(splans(11,nb),1,nz(nb),nq3(nb), ++ call sfftw_plan_many_dft(splans(11,nb),1,nz(nb),nq3(nb), + > scpl_mb(Atest(1)),nz(nb)*nq3(nb),1,nz(nb), + > scpl_mb(Atest(1)),nz(nb)*nq3(nb),1,nz(nb), + > ibackward,FFTW_EXHAUSTIVE) +- call fftw_plan_many_dft(splans(12,nb),1,ny(nb),nq2(nb), ++ call sfftw_plan_many_dft(splans(12,nb),1,ny(nb),nq2(nb), + > scpl_mb(Atest(1)),ny(nb)*nq2(nb),1,ny(nb), + > scpl_mb(Atest(1)),ny(nb)*nq2(nb),1,ny(nb), + > ibackward,FFTW_EXHAUSTIVE) +- call fftw_plan_many_dft_c2r(splans(13,nb),1,nx(nb), ++ call sfftw_plan_many_dft_c2r(splans(13,nb),1,nx(nb), + > nq1(nb), + > scpl_mb(Atest(1)),nfft3d(nb),1,nxh, + > scpl_mb(Atest(1)),n2ft3d(nb),1,(nx(nb)+2), + > FFTW_EXHAUSTIVE) +- call fftw_plan_many_dft_r2c(splans(14,nb),1,nx(nb), ++ call sfftw_plan_many_dft_r2c(splans(14,nb),1,nx(nb), + > nq1(nb), + > scpl_mb(Atest(1)),n2ft3d(nb),1,nx(nb)+2, + > scpl_mb(Atest(1)),nfft3d(nb),1,nxh, + > FFTW_EXHAUSTIVE) + +- call fftw_plan_many_dft(splans(15,nb),1,ny(nb),nq2(nb), ++ call sfftw_plan_many_dft(splans(15,nb),1,ny(nb),nq2(nb), + > scpl_mb(Atest(1)),ny(nb)*nq2(nb),1,ny(nb), + > scpl_mb(Atest(1)),ny(nb)*nq2(nb),1,ny(nb), + > iforward,FFTW_EXHAUSTIVE) +- call fftw_plan_many_dft(splans(16,nb),1,nz(nb),nq3(nb), ++ call sfftw_plan_many_dft(splans(16,nb),1,nz(nb),nq3(nb), + > scpl_mb(Atest(1)),nz(nb)*nq3(nb),1,nz(nb), + > scpl_mb(Atest(1)),nz(nb)*nq3(nb),1,nz(nb), + > iforward,FFTW_EXHAUSTIVE) + +- call fftw_plan_many_dft(splans(17,nb),1,nz(nb),1, ++ call sfftw_plan_many_dft(splans(17,nb),1,nz(nb),1, + > scpl_mb(Atest(1)),nz(nb),1,1, + > scpl_mb(Atest(1)),nz(nb),1,1, + > ibackward,FFTW_EXHAUSTIVE) +- call fftw_plan_many_dft(splans(18,nb),1,ny(nb),1, ++ call sfftw_plan_many_dft(splans(18,nb),1,ny(nb),1, + > scpl_mb(Atest(1)),ny(nb),1,1, + > scpl_mb(Atest(1)),ny(nb),1,1, + > ibackward,FFTW_EXHAUSTIVE) + +- call fftw_plan_many_dft(splans(19,nb),1,ny(nb),1, ++ call sfftw_plan_many_dft(splans(19,nb),1,ny(nb),1, + > scpl_mb(Atest(1)),ny(nb),1,1, + > scpl_mb(Atest(1)),ny(nb),1,1, + > iforward,FFTW_EXHAUSTIVE) +- call fftw_plan_many_dft(splans(20,nb),1,nz(nb),1, ++ call sfftw_plan_many_dft(splans(20,nb),1,nz(nb),1, + > scpl_mb(Atest(1)),nz(nb),1,1, + > scpl_mb(Atest(1)),nz(nb),1,1, + > iforward,FFTW_EXHAUSTIVE) +@@ -388,7 +389,7 @@ subroutine D3dBs_fft_end(nb) + + #ifdef FFTW3 + do i=1,nplans +- if (splans(i,nb).ne.0) call fftw_destroy_plan(splans(i,nb)) ++ if (splans(i,nb).ne.0) call sfftw_destroy_plan(splans(i,nb)) + end do + #endif + +@@ -497,7 +498,7 @@ subroutine D3dBs_cr_fft3b(nb,A) + #ifdef FFTW3 + do q=1,nq(nb) + indx = 1+(q-1)*nxhz +- call fftw_execute_dft(pslans(1,nb),A(indx),A(indx)) ++ call sfftw_execute_dft(splans(1,nb),A(indx),A(indx)) + end do + #else + indx0 = 0 +@@ -536,7 +537,7 @@ subroutine D3dBs_cr_fft3b(nb,A) + #ifdef FFTW3 + do q=1,nq(nb) + indx = 1+(q-1)*nxhy +- call fftw_execute_dft(splans(2,nb),A(indx),A(indx)) ++ call sfftw_execute_dft(splans(2,nb),A(indx),A(indx)) + end do + #else + indx0 = 0 +@@ -566,7 +567,7 @@ subroutine D3dBs_cr_fft3b(nb,A) + * ********************************************************************* + + #ifdef FFTW3 +- call fftw_execute_dft_c2r(splans(3,nb),A,A) ++ call sfftw_execute_dft_c2r(splans(3,nb),A,A) + + #else + +@@ -595,7 +596,7 @@ subroutine D3dBs_cr_fft3b(nb,A) + * ***************************************************** + + #ifdef FFTW3 +- call fftw_execute_dft(splans(11,nb),A,A) ++ call sfftw_execute_dft(splans(11,nb),A,A) + + #else + +@@ -615,7 +616,7 @@ subroutine D3dBs_cr_fft3b(nb,A) + * *** A(ny(nb),nz(nb),kx) <- fft1d^(-1)[A(ky,nz(nb),kx)] *** + * ************************************************************* + #ifdef FFTW3 +- call fftw_execute_dft(splans(12,nb),A,A) ++ call sfftw_execute_dft(splans(12,nb),A,A) + + #else + offset=tid*(2*ny(nb)+15) +@@ -634,7 +635,7 @@ subroutine D3dBs_cr_fft3b(nb,A) + * ********************************************************************* + + #ifdef FFTW3 +- call dfftw_execute_dft_c2r(splans(13,nb),A,A) ++ call sfftw_execute_dft_c2r(splans(13,nb),A,A) + + #else + offset=tid*(2*nx(nb)+15) +@@ -842,7 +843,7 @@ subroutine D3dBs_rc_fft3f(nb,A) + * *** A(kx,ny(nb),nz(nb)) <- fft1d[A(nx(nb),ny(nb),nz(nb))] *** + * **************************************************************** + #ifdef FFTW3 +- call dfftw_execute_dft_r2c(splans(4,nb),A,A) ++ call sfftw_execute_dft_r2c(splans(4,nb),A,A) + + #else + offset=tid*(2*nx(nb)+15) +@@ -860,7 +861,7 @@ subroutine D3dBs_rc_fft3f(nb,A) + #ifdef FFTW3 + do q=1,nq(nb) + indx = 1+(q-1)*nxhy +- call fftw_execute_dft(splans(5,nb),A(indx),A(indx)) ++ call sfftw_execute_dft(splans(5,nb),A(indx),A(indx)) + end do + + #else +@@ -902,7 +903,7 @@ subroutine D3dBs_rc_fft3f(nb,A) + #ifdef FFTW3 + do q=1,nq(nb) + indx = 1+(q-1)*nxhz +- call fftw_execute_dft(splans(6,nb),A(indx),A(indx)) ++ call sfftw_execute_dft(splans(6,nb),A(indx),A(indx)) + end do + + #else +@@ -947,7 +948,7 @@ subroutine D3dBs_rc_fft3f(nb,A) + * *** A(kx,ny(nb),nz(nb)) <- fft1d[A(nx(nb),ny(nb),nz(nb))] *** + * **************************************************************** + #ifdef FFTW3 +- call fftw_execute_dft_r2c(splans(14,nb),A,A) ++ call sfftw_execute_dft_r2c(splans(14,nb),A,A) + #else + offset=tid*(2*nx(nb)+15) + do i=tid+1,nq1(nb),nthr +@@ -964,7 +965,7 @@ subroutine D3dBs_rc_fft3f(nb,A) + * *** A(ky,nz(nb),kx) <- fft1d[A(ny(nb),nz(nb),kx)] *** + * ******************************************** + #ifdef FFTW3 +- call fftw_execute_dft(splans(15,nb),A,A) ++ call sfftw_execute_dft(splans(15,nb),A,A) + + #else + offset=tid*(2*ny(nb)+15) +@@ -982,7 +983,7 @@ subroutine D3dBs_rc_fft3f(nb,A) + * *** A(kz,kx,ky) <- fft1d[A(nz(nb),kx,ky)] *** + * ************************************************ + #ifdef FFTW3 +- call fftw_execute_dft(splans(16,nb),A,A) ++ call sfftw_execute_dft(splans(16,nb),A,A) + #else + offset=tid*(2*nz(nb)+15) + do i=tid+1,nq3(nb),nthr +diff --git a/src/nwpw/nwpwlib/D3dB/D3dBs_pfft.F b/src/nwpw/nwpwlib/D3dB/D3dBs_pfft.F +index f8fbd77bf6..bdbc5c91ba 100644 +--- a/src/nwpw/nwpwlib/D3dB/D3dBs_pfft.F ++++ b/src/nwpw/nwpwlib/D3dB/D3dBs_pfft.F +@@ -97,7 +97,7 @@ subroutine D3dBs_cr_pfft3b(nb,nbb,A) + + #ifdef FFTW3 + indx = i + indx0 +- call fftw_execute_dft(splans(7,nb),A(indx),A(indx)) ++ call sfftw_execute_dft(splans(7,nb),A(indx),A(indx)) + + #else + indx = i + indx0 +@@ -140,7 +140,7 @@ subroutine D3dBs_cr_pfft3b(nb,nbb,A) + + #ifdef FFTW3 + indx = i + indx0 +- call fftw_execute_dft(splans(8,nb),A(indx),A(indx)) ++ call sfftw_execute_dft(splans(8,nb),A(indx),A(indx)) + #else + indx = i + indx0 + indx1 = indx +@@ -166,7 +166,7 @@ subroutine D3dBs_cr_pfft3b(nb,nbb,A) + * ************************************************* + + #ifdef FFTW3 +- call fftw_execute_dft_c2r(splans(3,nb),A,A) ++ call sfftw_execute_dft_c2r(splans(3,nb),A,A) + + #else + call cshift1_sfftb(nx(nb),ny(nb),nq(nb),1,A) +@@ -197,7 +197,7 @@ subroutine D3dBs_cr_pfft3b(nb,nbb,A) + indx = 1 + do q=1,nq3(nb) + if (.not.log_mb(zero_row3(1,nbb)+q-1)) then +- call fftw_execute_dft(splans(17,nb),A(indx),A(indx)) ++ call sfftw_execute_dft(splans(17,nb),A(indx),A(indx)) + end if + indx = indx + nz(nb) + end do +@@ -229,7 +229,7 @@ subroutine D3dBs_cr_pfft3b(nb,nbb,A) + indx = 1 + do q=1,nq2(nb) + if (.not.log_mb(zero_row2(1,nbb)+q-1)) then +- call fftw_execute_dft(splans(18,nb),A(indx),A(indx)) ++ call sfftw_execute_dft(splans(18,nb),A(indx),A(indx)) + end if + indx = indx + ny(nb) + end do +@@ -258,7 +258,7 @@ subroutine D3dBs_cr_pfft3b(nb,nbb,A) + * ********************************************************************* + + #ifdef FFTW3 +- call fftw_execute_dft_c2r(splans(13,nb),A,A) ++ call sfftw_execute_dft_c2r(splans(13,nb),A,A) + #else + + !indx = 1 +@@ -368,7 +368,7 @@ subroutine D3dBs_rc_pfft3f(nb,nbb,A) + + + #ifdef FFTW3 +- call dfftw_execute_dft_r2c(splans(4,nb),A,A) ++ call sfftw_execute_dft_r2c(splans(4,nb),A,A) + + #else + indx = 1 +@@ -396,7 +396,7 @@ subroutine D3dBs_rc_pfft3f(nb,nbb,A) + + #ifdef FFTW3 + indx = i + indx0 +- call fftw_execute_dft(splans(9,nb),A(indx),A(indx)) ++ call sfftw_execute_dft(splans(9,nb),A(indx),A(indx)) + + #else + indx = i + indx0 +@@ -440,7 +440,7 @@ subroutine D3dBs_rc_pfft3f(nb,nbb,A) + + #ifdef FFTW3 + indx = i + indx0 +- call fftw_execute_dft(splans(10,nb),A(indx),A(indx)) ++ call sfftw_execute_dft(splans(10,nb),A(indx),A(indx)) + + #else + indx = i + indx0 +@@ -480,7 +480,7 @@ subroutine D3dBs_rc_pfft3f(nb,nbb,A) + * ******************************************** + + #ifdef FFTW3 +- call fftw_execute_dft_r2c(splans(14,nb),A,A) ++ call sfftw_execute_dft_r2c(splans(14,nb),A,A) + + #else + !indx = 1 +@@ -506,7 +506,7 @@ subroutine D3dBs_rc_pfft3f(nb,nbb,A) + indx = 1 + do q=1,nq2(nb) + if (.not.log_mb(zero_row2(1,nbb)+q-1)) then +- call fftw_execute_dft(splans(19,nb),A(indx),A(indx)) ++ call sfftw_execute_dft(splans(19,nb),A(indx),A(indx)) + end if + indx = indx + ny(nb) + end do +@@ -538,7 +538,7 @@ subroutine D3dBs_rc_pfft3f(nb,nbb,A) + indx = 1 + do q=1,nq3(nb) + if (.not.log_mb(zero_row3(1,nbb)+q-1)) then +- call fftw_execute_dft(splans(20,nb),A(indx),A(indx)) ++ call sfftw_execute_dft(splans(20,nb),A(indx),A(indx)) + end if + indx = indx + nz(nb) + end do +@@ -660,7 +660,7 @@ subroutine D3dBs_cr_mpfft3b(nb,nbb,m,A) + + #ifdef FFTW3 + indx = i+indx0+shift0 +- call fftw_execute_dft(splans(7,nb),A(indx),A(indx)) ++ call sfftw_execute_dft(splans(7,nb),A(indx),A(indx)) + + #else + indx = i + indx0 +@@ -710,7 +710,7 @@ subroutine D3dBs_cr_mpfft3b(nb,nbb,m,A) + + #ifdef FFTW3 + indx = i+indx0+shift0 +- call fftw_execute_dft(splans(8,nb),A(indx),A(indx)) ++ call sfftw_execute_dft(splans(8,nb),A(indx),A(indx)) + #else + indx = i + indx0 + indx1 = indx +@@ -736,7 +736,7 @@ subroutine D3dBs_cr_mpfft3b(nb,nbb,m,A) + * ********************************************************************* + + #ifdef FFTW3 +- call fftw_execute_dft_c2r(splans(3,nb),A(shift0),A(shift0)) ++ call sfftw_execute_dft_c2r(splans(3,nb),A(shift0),A(shift0)) + + #else + call cshift1_sfftb(nx(nb),ny(nb),nq(nb),1,A(shift)) +@@ -774,7 +774,7 @@ subroutine D3dBs_cr_mpfft3b(nb,nbb,m,A) + if (.not.log_mb(zero_row3(1,nbb)+q-1)) then + + #ifdef FFTW3 +- call fftw_execute_dft(splans(17,nb), ++ call sfftw_execute_dft(splans(17,nb), + > A(indx+shift0),A(indx+shift0)) + #else + call scfftb(nz(nb),A(shift0+indx),scpl_mb(tmpz(1,nb))) +@@ -798,7 +798,7 @@ subroutine D3dBs_cr_mpfft3b(nb,nbb,m,A) + if (.not.log_mb(zero_row2(1,nbb)+q-1)) then + + #ifdef FFTW3 +- call fftw_execute_dft(splans(18,nb), ++ call sfftw_execute_dft(splans(18,nb), + > A(indx+shift0),A(indx+shift0)) + #else + call scfftb(ny(nb),A(shift0+indx),scpl_mb(tmpy(1,nb))) +@@ -818,7 +818,7 @@ subroutine D3dBs_cr_mpfft3b(nb,nbb,m,A) + * ********************************************************************* + + #ifdef FFTW3 +- call fftw_execute_dft_c2r(splans(13,nb),A(shift0),A(shift0)) ++ call sfftw_execute_dft_c2r(splans(13,nb),A(shift0),A(shift0)) + + #else + call cshift1_sfftb(nx(nb),nq1(nb),1,1,A(shift)) +@@ -903,7 +903,7 @@ subroutine D3dBs_pfftfx(nbb,A,tmp1,tmp2,request,reqcnt) + call nwpw_timing_start(31) + + #ifdef FFTW3 +- call dfftw_execute_dft_r2c(splans(4,id),A,A) ++ call sfftw_execute_dft_r2c(splans(4,id),A,A) + + #else + +@@ -937,7 +937,7 @@ subroutine D3dBs_pfftfx(nbb,A,tmp1,tmp2,request,reqcnt) + + + #ifdef FFTW3 +- call fftw_execute_dft_r2c(splans(14,id),A,A) ++ call sfftw_execute_dft_r2c(splans(14,id),A,A) + + #else + +@@ -1030,7 +1030,7 @@ subroutine D3dBs_pfftfx0(nbb,tmp1,tmp2,request,reqcnt) + call nwpw_timing_start(31) + + #ifdef FFTW3 +- call fftw_execute_dft_r2c(splans(4,id),tmp1,tmp1) ++ call sfftw_execute_dft_r2c(splans(4,id),tmp1,tmp1) + + #else + +@@ -1059,7 +1059,7 @@ subroutine D3dBs_pfftfx0(nbb,tmp1,tmp2,request,reqcnt) + call nwpw_timing_start(31) + + #ifdef FFTW3 +- call fftw_execute_dft_r2c(splans(14,id),tmp1,tmp1) ++ call sfftw_execute_dft_r2c(splans(14,id),tmp1,tmp1) + + #else + +@@ -1154,7 +1154,7 @@ subroutine D3dBs_pfftfy(nbb,tmp1,tmp2,request,reqcnt) + indx2 = indx2 + 1 + if (.not.log_mb(zero_row2(1,nbb)+indx2-1)) then + indx = i + indx0 +- call fftw_execute_dft(splans(9,1),tmp1(indx),tmp1(indx)) ++ call sfftw_execute_dft(splans(9,1),tmp1(indx),tmp1(indx)) + end if + end do + indx0 = indx0 + nxhy +@@ -1214,7 +1214,7 @@ subroutine D3dBs_pfftfy(nbb,tmp1,tmp2,request,reqcnt) + indx = 1 + do q=1,nq2(id) + if (.not.log_mb(zero_row2(1,nbb)+q-1)) then +- call fftw_execute_dft(splans(19,1),tmp1(indx),tmp1(indx)) ++ call sfftw_execute_dft(splans(19,1),tmp1(indx),tmp1(indx)) + end if + indx = indx + ny(id) + end do +@@ -1333,7 +1333,7 @@ subroutine D3dBs_pfftfz(nbb,tmp1,tmp2,request,reqcnt) + + #ifdef FFTW3 + indx = i + indx0 +- call fftw_execute_dft(splans(10,1),tmp2(indx),tmp2(indx)) ++ call sfftw_execute_dft(splans(10,1),tmp2(indx),tmp2(indx)) + + #else + indx = i + indx0 +@@ -1374,7 +1374,7 @@ subroutine D3dBs_pfftfz(nbb,tmp1,tmp2,request,reqcnt) + indx = 1 + do q=1,nq3(id) + if (.not.log_mb(zero_row3(1,nbb)+q-1)) then +- call fftw_execute_dft(splans(20,1),tmp2(indx),tmp2(indx)) ++ call sfftw_execute_dft(splans(20,1),tmp2(indx),tmp2(indx)) + end if + indx = indx + nz(id) + end do +@@ -1605,7 +1605,7 @@ subroutine D3dBs_pfftbz(nbb,tmp1,tmp2,request,reqcnt) + + #ifdef FFTW3 + indx = i + indx0 +- call fftw_execute_dft(splans(7,1),tmp1(indx),tmp1(indx)) ++ call sfftw_execute_dft(splans(7,1),tmp1(indx),tmp1(indx)) + #else + + indx = i + indx0 +@@ -1652,7 +1652,7 @@ subroutine D3dBs_pfftbz(nbb,tmp1,tmp2,request,reqcnt) + indx = 1 + do q=1,nq3(id) + if (.not.log_mb(zero_row3(1,nbb)+q-1)) then +- call fftw_execute_dft(splans(17,1),tmp1(indx),tmp1(indx)) ++ call sfftw_execute_dft(splans(17,1),tmp1(indx),tmp1(indx)) + end if + indx = indx + nz(id) + end do +@@ -1768,7 +1768,7 @@ subroutine D3dBs_pfftby(nbb,tmp1,tmp2,request,reqcnt) + + #ifdef FFTW3 + indx = i + indx0 +- call fftw_execute_dft(splans(8,id),tmp2(indx),tmp2(indx)) ++ call sfftw_execute_dft(splans(8,id),tmp2(indx),tmp2(indx)) + + #else + indx = i + indx0 +@@ -1811,7 +1811,7 @@ subroutine D3dBs_pfftby(nbb,tmp1,tmp2,request,reqcnt) + indx = 1 + do q=1,nq2(id) + if (.not.log_mb(zero_row2(1,nbb)+q-1)) then +- call fftw_execute_dft(splans(18,id),tmp2(indx),tmp2(indx)) ++ call sfftw_execute_dft(splans(18,id),tmp2(indx),tmp2(indx)) + end if + indx = indx + ny(id) + end do +@@ -1912,7 +1912,7 @@ subroutine D3dBs_pfftbx(nbb,tmp1,tmp2,request,reqcnt) + + + #ifdef FFTW3 +- call fftw_execute_dft_c2r(splans(3,id),tmp2,tmp2) ++ call sfftw_execute_dft_c2r(splans(3,id),tmp2,tmp2) + + #else + #ifdef NUSE_OPENMP +@@ -1956,7 +1956,7 @@ subroutine D3dBs_pfftbx(nbb,tmp1,tmp2,request,reqcnt) + * ********************************************************************* + + #ifdef FFTW3 +- call fftw_execute_dft_c2r(splans(13,id),tmp1,tmp1) ++ call sfftw_execute_dft_c2r(splans(13,id),tmp1,tmp1) + + #else + indx = 1 +diff --git a/src/nwpw/nwpwlib/D3dB/GNUmakefile b/src/nwpw/nwpwlib/D3dB/GNUmakefile +index df7b8b7068..ec1fa61c42 100644 +--- a/src/nwpw/nwpwlib/D3dB/GNUmakefile ++++ b/src/nwpw/nwpwlib/D3dB/GNUmakefile +@@ -52,5 +52,7 @@ endif + + ifeq ($(USE_FFTW3),y) + LIB_DEFINES += -DFFTW3 +- LIB_INCLUDES += -I$(FFTW3_INCLUDE) ++ ifdef FFTW3_INCLUDE ++ LIB_INCLUDES += -I$(FFTW3_INCLUDE) ++ endif + endif +From 376f86f96eb982e83f10514e9dcd994564f973b4 Mon Sep 17 00:00:00 2001 +From: Eric Bylaska +Date: Thu, 8 Jun 2023 15:41:50 -0700 +Subject: [PATCH] Fixed singlet precision fftw3 linking. In general, fftw3 + should not be used by most users, but it is included for perfmance testing + and cross checking...EJB + +--- + src/config/makefile.h | 2 +- + 1 file changed, 1 insertion(+), 1 deletion(-) + +diff --git a/src/config/makefile.h b/src/config/makefile.h +index 1f2ff5c472..8b2b667bd9 100644 +--- a/src/config/makefile.h ++++ b/src/config/makefile.h +@@ -3622,7 +3622,7 @@ endif + # FFTW3 library inclusion + ifdef USE_FFTW3 + ifndef LIBFFTW3 +- LIBFFTW3 = -lfftw3 ++ LIBFFTW3 = -lfftw3 -lfftw3f + endif + ifdef FFTW3_LIB + CORE_LIBS += -L$(FFTW3_LIB) +From c89fc9d1eca6689bce12564a63fdea95d962a123 Mon Sep 17 00:00:00 2001 +From: edoapra +Date: Thu, 8 Jun 2023 17:37:13 -0700 +Subject: [PATCH] fix for including system header fftw3.f + +--- + src/nwpw/nwpwlib/D3dB/D3dB-new.F | 3 +-- + src/nwpw/nwpwlib/D3dB/D3dBs-new.F | 3 +-- + 2 files changed, 2 insertions(+), 4 deletions(-) + +diff --git a/src/nwpw/nwpwlib/D3dB/D3dB-new.F b/src/nwpw/nwpwlib/D3dB/D3dB-new.F +index 2159707831..6249f91710 100644 +--- a/src/nwpw/nwpwlib/D3dB/D3dB-new.F ++++ b/src/nwpw/nwpwlib/D3dB/D3dB-new.F +@@ -1183,8 +1183,7 @@ subroutine D3dB_fft_init(nb) + common / D3dB_fft / tmpx,tmpy,tmpz + + #ifdef FFTW3 +-!#include "fftw3.fh" +-#include "fftw3.f" ++#include + integer Atest(2),nxh,nxhz,nxhy + #endif + +diff --git a/src/nwpw/nwpwlib/D3dB/D3dBs-new.F b/src/nwpw/nwpwlib/D3dB/D3dBs-new.F +index 8ddfddad34..4b71e4cf29 100644 +--- a/src/nwpw/nwpwlib/D3dB/D3dBs-new.F ++++ b/src/nwpw/nwpwlib/D3dB/D3dBs-new.F +@@ -195,8 +195,7 @@ subroutine D3dBs_fft_init(nb) + common / D3dBs_fft / tmpx,tmpy,tmpz + + #ifdef FFTW3 +-!#include "fftw3.fh" +-#include "fftw3.f" ++#include + integer Atest(2),nxh,nxhz,nxhy + #endif + diff --git a/var/spack/repos/builtin/packages/nwchem/package.py b/var/spack/repos/builtin/packages/nwchem/package.py index fb2aa60167..aeb55ab319 100644 --- a/var/spack/repos/builtin/packages/nwchem/package.py +++ b/var/spack/repos/builtin/packages/nwchem/package.py @@ -30,12 +30,23 @@ class Nwchem(Package): variant("openmp", default=False, description="Enables OpenMP support") variant("mpipr", default=False, description="Enables ARMCI with progress rank") + variant("fftw3", default=False, description="Link against the FFTW library") # This patch is for the modification of the build system (e.g. compiler flags) and # Fortran syntax to enable the compilation with Fujitsu compilers. The modification # will be merged to the next release of NWChem (see https://github.com/nwchemgit/nwchem/issues/347 # for more detail. patch("fj.patch", when="@7.0.2 %fj") + # This patch is for linking the FFTW library in NWChem. + # It applys only to the 7.2.0 source code. + # will be merged to the next release of NWChem (see https://github.com/nwchemgit/nwchem/issues/792 + # for more detail. + # This patch is the combination of the following commits + # https://github.com/nwchemgit/nwchem/commit/b4ec4ade1af434bc80470d6874aebf6fdcd12489 + # https://github.com/nwchemgit/nwchem/commit/376f86f96eb982e83f10514e9dcd994564f973b4 + # https://github.com/nwchemgit/nwchem/commit/c89fc9d1eca6689bce12564a63fdea95d962a123 + # Prior versions of NWChem, including 7.0.2, were not able to link with FFTW + patch("fftw_splans.patch", when="@7.2.0") depends_on("blas") depends_on("lapack") -- cgit v1.2.3-60-g2f50