Merge pull request #2 from yukwangmin/master

PHonon - k point implementation

Author: fspiga

PH/Makefile

@@ -15,13 +15,24 @@ LIBOBJS        = ../../clib/clib.a ../../iotk/src/libiotk.a
 
 TLDEPS= bindir mods libs pw-gpu
 
+
+PHGPUOBJS = \
+h_psiq_gpu.o
+
+CUGPUOBJS = \
+h_psiq_k_cuda.o
+
 all : tldeps ph-gpu.x
 
-ph-gpu.x : phonon.o ../../PHonon/PH/libph.a $(PWMODS) $(LIBOBJS)
-	$(LD) $(LDFLAGS) -o $@ phonon.o ../../PHonon/PH/libph.a \
+ph-gpu.x : phonon.o ../../PHonon/PH/libph.a libphgpu.a $(PWMODS) $(LIBOBJS)
+	$(LD) $(LDFLAGS) -o $@ phonon.o ../../PHonon/PH/libph.a libphgpu.a \
 		 $(PWMODS) $(QEMODS) $(LIBOBJS) $(LIBS)
 	- ( cd ../../bin ; ln -fs ../GPU/PH/$@ . )
 
+libphgpu.a : $(PHGPUOBJS) $(CUGPUOBJS)
+	$(AR) $(ARFLAGS) $@ $?
+	$(RANLIB) $@
+
 tldeps:
 	test -n "$(TLDEPS)" && ( cd ../.. ; $(MAKE) $(MFLAGS) $(TLDEPS) || exit 1) || :
 clean :

PH/h_psiq_gpu.f90

@@ -0,0 +1,59 @@
+!
+! \author Kwanmgin Yu <[email protected]>
+!
+!
+
+
+!---------------------------
+SUBROUTINE h_psiq_k_gpu(lda, n, m, psi, hpsi, spsi)
+!---------------------------
+  USE kinds,   ONLY : DP
+  USE gvecs,   ONLY : nls, ngms
+  USE wvfct,  ONLY : g2kin, npwx
+  USE fft_base,      ONLY : dffts
+  USE wavefunctions_module,  ONLY : psic, psic_nc
+  USE becmod, ONLY : bec_type, becp, calbec
+  USE noncollin_module, ONLY : noncolin, npol
+  USE lsda_mod, ONLY : current_spin
+  USE fft_interfaces, ONLY: fwfft, invfft
+  USE spin_orb, ONLY : domag
+  USE scf,    ONLY : vrs
+  USE uspp,   ONLY : vkb
+  USE qpoint, ONLY: igkq, nksq
+  USE control_flags, ONLY : gamma_only ! Needed only for TDDFPT
+  USE mp_pools,  ONLY: npool
+
+
+  IMPLICIT NONE
+  !
+  INTEGER, INTENT(in) :: lda, n, m
+  !INTEGER, POINTER :: igkq(:)
+  COMPLEX(DP), INTENT(INOUT)  :: psi (lda*npol, m)
+  COMPLEX(DP), INTENT(OUT) :: hpsi (lda*npol, m), spsi (lda*npol, m)
+  !
+  INTEGER, EXTERNAL :: h_psiq_cuda_k
+  !
+  INTEGER :: ierr
+  INTEGER :: i,j
+  
+  
+  
+  CALL calbec ( n, vkb, psi, becp, m)
+
+  ierr = h_psiq_cuda_k( lda, npol, dffts%nnr, dffts%nr1x, dffts%nr2x, dffts%nr3x, n, m, psi, vrs(:,current_spin), hpsi, igkq(1:), nls(1:), ngms, g2kin)
+  
+
+  IF (ierr.EQ.1) THEN
+    !CPU fall-back
+    write (*,*) 'h_psiq_k_gpu()::ierr : ', ierr
+    CALL h_psiq()
+  ENDIF
+  
+  CALL add_vuspsi (lda, n, m, hpsi)
+
+  CALL s_psi (lda, n, m, psi, spsi)
+  
+  !
+  RETURN
+END SUBROUTINE h_psiq_k_gpu
+