hsmt_ab.F90 9.11 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11
!--------------------------------------------------------------------------------
! Copyright (c) 2016 Peter Grünberg Institut, Forschungszentrum Jülich, Germany
! This file is part of FLEUR and available as free software under the conditions
! of the MIT license as expressed in the LICENSE file in more detail.
!--------------------------------------------------------------------------------
MODULE m_hsmt_ab
  use m_juDFT
  implicit none

  INTERFACE hsmt_ab
    module procedure hsmt_ab_cpu
12
#ifdef _CUDA
13 14 15 16 17 18 19
    module procedure hsmt_ab_gpu
#endif
  END INTERFACE


CONTAINS

20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56
#ifdef _CUDA

  ATTRIBUTES(global) SUBROUTINE synth_ab(grid,block,n,lmax,iintsp,ab_size,gkrot_dev,fj,gj,c_ph,ab)
    USE m_ylm
    INTEGER, VALUE, INTENT(IN) :: grid, block, n, lmax, iintsp,ab_size
    REAL,   DEVICE, INTENT(IN) :: gkrot_dev(:,:),fj(:,:,:),gj(:,:,:)
    COMPLEX,DEVICE, INTENT(IN) :: c_ph(:,:)
    COMPLEX,DEVICE, INTENT (OUT) :: ab(:,:)
    COMPLEX,ALLOCATABLE :: ylm(:)
    INTEGER :: k,l,ll1,m
    INTEGER :: loop_start, loop_end, i, loop_size

    ALLOCATE(ylm((lmax+1)**2))

    k = (blockidx%x-1)*blockdim%x + threadidx%x

    loop_size = max(n/(grid*block),1)
    if (loop_size * grid*block < n) loop_size = loop_size + 1
    loop_start = (k-1) * loop_size + 1
    loop_end = loop_start + loop_size - 1
    if (loop_end > n ) loop_end = n

    DO i = loop_start,loop_end
       !-->    generate spherical harmonics
       CALL ylm4_dev(lmax,gkrot_dev(:,i),ylm(:))
       DO l = 0,lmax
          ll1 = l* (l+1)
          DO m = -l,l               
             ab(i,ll1+m+1)         = CONJG(fj(i,l+1,iintsp)*c_ph(i,iintsp)*ylm(ll1+m+1)) 
             ab(i,ll1+m+1+ab_size) = CONJG(gj(i,l+1,iintsp)*c_ph(i,iintsp)*ylm(ll1+m+1)) 
          END DO
       END DO
    ENDDO 

    DEALLOCATE(ylm)
  END SUBROUTINE synth_ab

57 58

  SUBROUTINE hsmt_ab_gpu(sym,atoms,noco,ispin,iintsp,n,na,cell,lapw,fj,gj,ab,ab_size,l_nonsph,abclo,alo1,blo1,clo1)
59
!Calculate overlap matrix, GPU version
60 61 62 63
    USE m_constants, ONLY : fpi_const,tpi_const
    USE m_types
    USE m_ylm
    USE m_apws
64 65
    USE cudafor
    USE nvtx
66 67 68 69 70 71 72 73 74 75 76 77 78
    IMPLICIT NONE
    TYPE(t_sym),INTENT(IN)      :: sym
    TYPE(t_cell),INTENT(IN)     :: cell
    TYPE(t_atoms),INTENT(IN)    :: atoms
    TYPE(t_lapw),INTENT(IN)     :: lapw
    TYPE(t_noco),INTENT(IN)     :: noco
    !     ..
    !     .. Scalar Arguments ..
    INTEGER, INTENT (IN) :: ispin,n,na,iintsp
    LOGICAL,INTENT(IN)   :: l_nonsph
    INTEGER,INTENT(OUT)  :: ab_size
    !     ..
    !     .. Array Arguments ..
79
    REAL, DEVICE, INTENT(IN)       :: fj(:,:,:),gj(:,:,:)
80 81 82 83 84 85
    COMPLEX,DEVICE, INTENT (OUT) :: ab(:,:)
    !Optional arguments if abc coef for LOs are needed
    COMPLEX, INTENT(INOUT),OPTIONAL:: abclo(:,-atoms%llod:,:,:)
    REAL,INTENT(IN),OPTIONAL:: alo1(:),blo1(:),clo1(:)
    
    INTEGER:: np,k,l,ll1,m,lmax,nkvec,lo,lm,invsfct
Uliana Alekseeva's avatar
Uliana Alekseeva committed
86 87 88 89 90 91 92
    REAL   :: th,v(3),bmrot(3,3),vmult(3)
    COMPLEX,ALLOCATABLE :: ylm(:,:)
    COMPLEX,ALLOCATABLE :: c_ph(:,:)
    REAL,   ALLOCATABLE :: gkrot(:,:)
    COMPLEX:: term

    COMPLEX,ALLOCATABLE,DEVICE :: c_ph_dev(:,:)
93
    REAL,   ALLOCATABLE,DEVICE :: gkrot_dev(:,:)
94
    INTEGER :: istat, grid, block
Uliana Alekseeva's avatar
Uliana Alekseeva committed
95
 
96
    !call nvtxStartRange("hsmt_ab",2)    
97 98
    lmax=MERGE(atoms%lnonsph(n),atoms%lmax(n),l_nonsph)

Uliana Alekseeva's avatar
Uliana Alekseeva committed
99
    ALLOCATE(c_ph_dev(lapw%nv(1),MERGE(2,1,noco%l_ss)))
100
    ALLOCATE(gkrot_dev(3,lapw%nv(1)))
101

102
    ALLOCATE(ylm((lmax+1)**2,lapw%nv(1)))
103 104 105 106 107 108 109 110 111 112
    ALLOCATE(c_ph(lapw%nv(1),MERGE(2,1,noco%l_ss)))
    ALLOCATE(gkrot(3,lapw%nv(1)))

    
    ab_size=lmax*(lmax+2)+1
    ab=0.0
    
    np = sym%invtab(atoms%ngopr(na))
    !--->          set up phase factors
    CALL lapw%phase_factors(iintsp,atoms%taual(:,na),noco%qss,c_ph(:,iintsp))
Uliana Alekseeva's avatar
Uliana Alekseeva committed
113 114
    c_ph_dev=c_ph   
 
115 116 117 118 119 120 121 122 123 124 125 126 127
    IF (np==1) THEN
       gkrot(:, 1:lapw%nv(iintsp)) = lapw%gk(:, 1:lapw%nv(iintsp),iintsp)
    ELSE
       bmrot=MATMUL(1.*sym%mrot(:,:,np),cell%bmat)
       DO k = 1,lapw%nv(iintsp)
          !-->  apply the rotation that brings this atom into the
          !-->  representative (this is the definition of ngopr(na)
          !-->  and transform to cartesian coordinates
          v(:) = lapw%vk(:,k,iintsp)
          gkrot(:,k) = MATMUL(TRANSPOSE(bmrot),v)
       END DO
    END IF

128 129 130 131
    gkrot_dev = gkrot 


    !-->  synthesize the complex conjugates of a and b
132 133
    !call nvtxStartRange("hsmt_synthAB",5)    
    istat = cudaDeviceSynchronize() 
134

135 136 137 138
    ! pretty ugly solution
    block = 256
    grid = lapw%nv(1)/(block*4) + 1
    CALL synth_ab<<<grid,block>>>(grid,block,lapw%nv(1),lmax,iintsp,ab_size,gkrot_dev,fj,gj,c_ph_dev,ab)
139

140
    istat = cudaDeviceSynchronize() 
141
    !call nvtxEndRange
142 143

    IF (PRESENT(abclo)) THEN
144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163
       print*, "Ooooops, TODO in hsmt_ab"
       !DO k = 1,lapw%nv(1)
       !   !determine also the abc coeffs for LOs
       !   invsfct=MERGE(1,2,atoms%invsat(na).EQ.0)
       !   term = fpi_const/SQRT(cell%omtil)* ((atoms%rmt(n)**2)/2)*c_ph(k,iintsp)
       !   DO lo = 1,atoms%nlo(n)
       !      l = atoms%llo(lo,n)
       !      DO nkvec=1,invsfct*(2*l+1)
       !         IF (lapw%kvec(nkvec,lo,na)==k) THEN !This k-vector is used in LO
       !            ll1 = l*(l+1) + 1
       !            DO m = -l,l
       !               lm = ll1 + m
       !               abclo(1,m,nkvec,lo) = term*ylm(k,lm)*alo1(lo)
       !               abclo(2,m,nkvec,lo) = term*ylm(k,lm)*blo1(lo)
       !               abclo(3,m,nkvec,lo) = term*ylm(k,lm)*clo1(lo)
       !            END DO
       !         END IF
       !      ENDDO
       !   ENDDO
       !ENDDO
164 165
    ENDIF
       
166
    ab_size=ab_size*2
167
    
168 169 170
    DEALLOCATE(c_ph_dev)
    DEALLOCATE(gkrot_dev)

171 172 173 174
  END SUBROUTINE hsmt_ab_gpu
#endif

  SUBROUTINE hsmt_ab_cpu(sym,atoms,noco,ispin,iintsp,n,na,cell,lapw,fj,gj,ab,ab_size,l_nonsph,abclo,alo1,blo1,clo1)
175
!Calculate overlap matrix, CPU vesion
176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199
    USE m_constants, ONLY : fpi_const,tpi_const
    USE m_types
    USE m_ylm
    USE m_apws
    IMPLICIT NONE
    TYPE(t_sym),INTENT(IN)      :: sym
    TYPE(t_cell),INTENT(IN)     :: cell
    TYPE(t_atoms),INTENT(IN)    :: atoms
    TYPE(t_lapw),INTENT(IN)     :: lapw
    TYPE(t_noco),INTENT(IN)     :: noco
    !     ..
    !     .. Scalar Arguments ..
    INTEGER, INTENT (IN) :: ispin,n,na,iintsp
    LOGICAL,INTENT(IN)   :: l_nonsph
    INTEGER,INTENT(OUT)  :: ab_size
    !     ..
    !     .. Array Arguments ..
    REAL,INTENT(IN)       :: fj(:,0:,:),gj(:,0:,:)
    COMPLEX, INTENT (OUT) :: ab(:,:)
    !Optional arguments if abc coef for LOs are needed
    COMPLEX, INTENT(INOUT),OPTIONAL:: abclo(:,-atoms%llod:,:,:)
    REAL,INTENT(IN),OPTIONAL:: alo1(:),blo1(:),clo1(:)
    
    INTEGER:: np,k,l,ll1,m,lmax,nkvec,lo,lm,invsfct
Uliana Alekseeva's avatar
Uliana Alekseeva committed
200 201
    COMPLEX:: term
    REAL   :: th,v(3),bmrot(3,3),vmult(3)
202
    COMPLEX :: ylm((atoms%lmaxd+1)**2)
Uliana Alekseeva's avatar
Uliana Alekseeva committed
203 204
    COMPLEX,ALLOCATABLE:: c_ph(:,:)
    REAL,ALLOCATABLE   :: gkrot(:,:)
205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274
    LOGICAL :: l_apw
   
    ALLOCATE(c_ph(lapw%nv(1),MERGE(2,1,noco%l_ss)))
    ALLOCATE(gkrot(3,lapw%nv(1)))

    lmax=MERGE(atoms%lnonsph(n),atoms%lmax(n),l_nonsph)
    
    ab_size=lmax*(lmax+2)+1
    l_apw=ALL(gj==0.0)
    ab=0.0
    
    np = sym%invtab(atoms%ngopr(na))
    !--->          set up phase factors
    CALL lapw%phase_factors(iintsp,atoms%taual(:,na),noco%qss,c_ph(:,iintsp))
    
    IF (np==1) THEN
       gkrot(:, 1:lapw%nv(iintsp)) = lapw%gk(:, 1:lapw%nv(iintsp),iintsp)
    ELSE
       bmrot=MATMUL(1.*sym%mrot(:,:,np),cell%bmat)
       DO k = 1,lapw%nv(iintsp)
          !-->  apply the rotation that brings this atom into the
          !-->  representative (this is the definition of ngopr(na)
          !-->  and transform to cartesian coordinates
          v(:) = lapw%vk(:,k,iintsp)
          gkrot(:,k) = MATMUL(TRANSPOSE(bmrot),v)
       END DO
    END IF
    !$OMP PARALLEL DO DEFAULT(none) &
    !$OMP& SHARED(lapw,gkrot,lmax,c_ph,iintsp,ab,fj,gj,abclo,cell,atoms) &
    !$OMP& SHARED(alo1,blo1,clo1,ab_size,na,n) &
    !$OMP& PRIVATE(k,vmult,ylm,l,ll1,m,lm,term,invsfct,lo,nkvec)
    DO k = 1,lapw%nv(1)
       !-->    generate spherical harmonics
       vmult(:) =  gkrot(:,k)
       CALL ylm4(lmax,vmult,ylm)
       !-->  synthesize the complex conjugates of a and b
       DO l = 0,lmax
          ll1 = l* (l+1)
          DO m = -l,l               
             term = c_ph(k,iintsp)*ylm(ll1+m+1)
             ab(k,ll1+m+1)         = fj(k,l,iintsp)*term
             ab(k,ll1+m+1+ab_size) = gj(k,l,iintsp)*term
          END DO
       END DO
       IF (PRESENT(abclo)) THEN
          !determine also the abc coeffs for LOs
          invsfct=MERGE(1,2,atoms%invsat(na).EQ.0)
          term = fpi_const/SQRT(cell%omtil)* ((atoms%rmt(n)**2)/2)*c_ph(k,iintsp)
          DO lo = 1,atoms%nlo(n)
             l = atoms%llo(lo,n)
             DO nkvec=1,invsfct*(2*l+1)
                IF (lapw%kvec(nkvec,lo,na)==k) THEN !This k-vector is used in LO
                   ll1 = l*(l+1) + 1
                   DO m = -l,l
                      lm = ll1 + m
                      abclo(1,m,nkvec,lo) = term*ylm(lm)*alo1(lo)
                      abclo(2,m,nkvec,lo) = term*ylm(lm)*blo1(lo)
                      abclo(3,m,nkvec,lo) = term*ylm(lm)*clo1(lo)
                   END DO
                END IF
             ENDDO
          ENDDO
       ENDIF
       
    ENDDO !k-loop
    !$OMP END PARALLEL DO
    IF (.NOT.l_apw) ab_size=ab_size*2
    
  END SUBROUTINE hsmt_ab_cpu
END MODULE m_hsmt_ab