Removing assembly files

Compilation of AVX code with ICC is is now possible

README-AVX512

@@ -9,8 +9,6 @@ DESCRIPTION
 
    The extension is enabled with the -DAVX521 flag in CFLAGS in the Makefiles.
 
-   To enable assembly versions pre-compiled for Skylake Xeon Scalable CPUs also include -DAVX512_ASM in CFLAGS.
-
 
 
 AUTHORS

devel/dirac/Makefile

@@ -89,7 +89,7 @@ LIBPATH = $(MPI_HOME)/lib
 
 CFLAGS = -std=c89 -pedantic -fstrict-aliasing \
          -Wall -Wno-long-long -Wstrict-prototypes \
-         -O2 -DAVX512
+         -O2 -DAVX512 -DAVX -DPM
 
 LFLAGS = $(CFLAGS)
 

devel/linalg/Makefile

@@ -25,7 +25,7 @@ FLAGS = flags lat_parms dfl_parms
 
 LATTICE = bcnds uidx geometry
 
-LINALG = cmatrix_dble liealg salg salg_dble valg valg_dble
+LINALG = cmatrix_dble liealg salg salg_dble valg valg_dble salg_avx512 salg_dble_avx512
 
 RANDOM = ranlux ranlxs ranlxd gauss
 
@@ -42,9 +42,6 @@ VFLDS = vflds vinit
 MODULES = $(FLAGS) $(LATTICE) $(RANDOM) $(LINALG) $(UTILS) \
           $(UFLDS) $(SFLDS) $(SU3FCTS) $(VFLDS)
 
-AVX512_MODULES = salg_avx512 salg_dble_avx512
-
-AVX512_ASM_MODULES = salg_avx512_asm salg_dble_avx512_asm
 
 # Logging option (-mpilog or -mpitrace or -mpianim)
 
@@ -76,7 +73,7 @@ LIBPATH = $(MPI_HOME)/lib
 
 CFLAGS = -std=c89 -pedantic -fstrict-aliasing \
          -Wall -Wno-long-long -Wstrict-prototypes \
-         -O2 -DAVX512
+         -O2 -DAVX512 -DAVX -DPM
 
 LFLAGS = $(CFLAGS)
 
@@ -87,17 +84,6 @@ SHELL=/bin/bash
 CC=mpicc
 CLINKER=$(CC)
 
-#Check CFLAGS to find which AVX512 flags are active
-ifneq (,$(findstring AVX512_ASM,$(CFLAGS)))
-MODULES= $(STD_MODULES) $(AVX512_ASM_MODULES)
-else
-ifneq (,$(findstring AVX512,$(CFLAGS)))
-MODULES= $(STD_MODULES) $(AVX512_MODULES)
-else
-MODULES= $(STD_MODULES)
-endif
-endif
-
 PGMS= $(MAIN) $(MODULES)
 
 -include $(addsuffix .d,$(PGMS))
@@ -109,14 +95,7 @@ $(addsuffix .d,$(PGMS)): %.d: %.c Makefile
 
 
 # rule to compile source programs
-$(addsuffix .o,$(MAIN) $(STD_MODULES)): %.o: %.c Makefile
-	$(CC) $< -c $(CFLAGS) $(LOGOPTION) $(addprefix -I,$(INCPATH))
-
-# pattern to compile files in the avx512 directiories
-$(addsuffix .o,$(AVX512_MODULES)): %.o: %.c Makefile
-	$(CC) $< -c $(CFLAGS) $(LOGOPTION) $(addprefix -I,$(INCPATH))
-
-$(addsuffix .o,$(AVX512_ASM_MODULES)): %.o: %.s Makefile
+$(addsuffix .o,$(MAIN) $(MODULES)): %.o: %.c Makefile
 	$(CC) $< -c $(CFLAGS) $(LOGOPTION) $(addprefix -I,$(INCPATH))
 
 # rule to link object files

devel/sw_term/Makefile

@@ -25,31 +25,25 @@ FLAGS = flags lat_parms dfl_parms
 
 LATTICE = bcnds uidx ftidx geometry
 
-LINALG = salg_dble liealg cmatrix_dble
+LINALG = salg_dble liealg cmatrix_dble salg_dble_avx512
 
 RANDOM = ranlux ranlxs ranlxd gauss
 
 SFLDS = sflds
 
 SU3FCTS = chexp su3prod su3ren cm3x3 random_su3
 
-SW_TERM = pauli pauli_dble swflds sw_term
+SW_TERM = pauli pauli_dble swflds sw_term pauli_avx512 pauli_dble_avx512
 
 TCHARGE = ftcom ftensor
 
 UFLDS = plaq_sum shift uflds udcom
 
 UTILS = endian error hsum mutils utils wspace
 
-STD_MODULES = $(FLAGS) $(LATTICE) $(LINALG) $(RANDOM) $(UFLDS) \
+MODULES = $(FLAGS) $(LATTICE) $(LINALG) $(RANDOM) $(UFLDS) \
           $(SU3FCTS) $(UTILS) $(SFLDS) $(TCHARGE) $(SW_TERM)
 
-AVX512_MODULES = salg_dble_avx512 \
-					pauli_avx512 pauli_dble_avx512
-
-AVX512_ASM_MODULES = salg_dble_avx512_asm \
- 					pauli_avx512_asm pauli_dble_avx512_asm
-
 # Logging option (-mpilog or -mpitrace or -mpianim)
 
 LOGOPTION =
@@ -62,7 +56,7 @@ MDIR = ../../modules
 VPATH = .:$(MDIR)/flags:$(MDIR)/lattice:$(MDIR)/linalg:$(MDIR)/random:\
           $(MDIR)/uflds:$(MDIR)/su3fcts:$(MDIR)/utils:$(MDIR)/sflds:\
           $(MDIR)/tcharge:$(MDIR)/sw_term:\
-					$(MDIR)/linalg/avx512:$(MDIR)/sw_term/avx512
+          $(MDIR)/linalg/avx512:$(MDIR)/sw_term/avx512
 
 
 # additional include directories
@@ -81,7 +75,7 @@ LIBPATH = $(MPI_HOME)/lib
 
 CFLAGS = -std=c89 -pedantic -fstrict-aliasing \
          -Wall -Wno-long-long -Wstrict-prototypes \
-         -O2 -DAVX512
+         -O2 -DAVX512 -DAVX -DPM
 
 LFLAGS = $(CFLAGS)
 
@@ -91,17 +85,6 @@ SHELL=/bin/bash
 CC=mpicc
 CLINKER=$(CC)
 
-#Check CFLAGS to find which AVX512 flags are active
-ifneq (,$(findstring AVX512_ASM,$(CFLAGS)))
-MODULES= $(STD_MODULES) $(AVX512_ASM_MODULES)
-else
-ifneq (,$(findstring AVX512,$(CFLAGS)))
-MODULES= $(STD_MODULES) $(AVX512_MODULES)
-else
-MODULES= $(STD_MODULES)
-endif
-endif
-
 PGMS= $(MAIN) $(MODULES)
 
 -include $(addsuffix .d,$(PGMS))
@@ -113,14 +96,7 @@ $(addsuffix .d,$(PGMS)): %.d: %.c Makefile
 
 
 # rule to compile source programs
-$(addsuffix .o,$(MAIN) $(STD_MODULES)): %.o: %.c Makefile
-	$(CC) $< -c $(CFLAGS) $(LOGOPTION) $(addprefix -I,$(INCPATH))
-
-# pattern to compile files in the avx512 directiories
-$(addsuffix .o,$(AVX512_MODULES)): %.o: %.c Makefile
-	$(CC) $< -c $(CFLAGS) $(LOGOPTION) $(addprefix -I,$(INCPATH))
-
-$(addsuffix .o,$(AVX512_ASM_MODULES)): %.o: %.s Makefile
+$(addsuffix .o,$(MAIN) $(MODULES)): %.o: %.c Makefile
 	$(CC) $< -c $(CFLAGS) $(LOGOPTION) $(addprefix -I,$(INCPATH))
 
 # rule to link object files

main/Makefile

@@ -26,7 +26,7 @@ BLOCK = block blk_grid map_u2blk map_sw2blk map_s2blk
 
 DFL = dfl_geometry dfl_subspace ltl_gcr dfl_sap_gcr dfl_modes
 
-DIRAC = Dw_dble Dw Dw_bnd
+DIRAC = Dw_dble Dw Dw_bnd Dw_avx512 Dw_dble_avx512
 
 FLAGS = flags action_parms dfl_parms force_parms hmc_parms lat_parms \
         mdint_parms rat_parms rw_parms sap_parms solver_parms
@@ -36,7 +36,8 @@ FORCES = force0 force1 force2 force3 force4 force5 \
 
 LATTICE = bcnds uidx ftidx geometry
 
-LINALG = salg salg_dble valg valg_dble liealg cmatrix_dble cmatrix
+LINALG = salg salg_dble valg valg_dble liealg cmatrix_dble cmatrix \
+         salg_avx512 salg_dble_avx512
 
 LINSOLV = cgne mscg fgcr fgcr4vd
 
@@ -54,7 +55,7 @@ SFLDS = sflds scom sdcom Pbnd Pbnd_dble
 
 SU3FCTS = chexp su3prod su3ren cm3x3 random_su3
 
-SW_TERM = pauli pauli_dble swflds sw_term
+SW_TERM = pauli pauli_dble swflds sw_term pauli_avx512 pauli_dble_avx512
 
 TCHARGE = ftcom ftensor tcharge ym_action
 
@@ -74,12 +75,6 @@ STD_MODULES = $(ARCHIVE) $(BLOCK) $(DFL) $(DIRAC) $(FLAGS) $(FORCES) \
           $(UFLDS) $(UPDATE) $(UTILS) $(VFLDS) $(WFLOW)
 
 
-AVX512_MODULES = Dw_avx512 Dw_dble_avx512 salg_avx512 salg_dble_avx512 \
-	pauli_avx512 pauli_dble_avx512
-
-AVX512_ASM_MODULES = Dw_avx512_asm Dw_dble_avx512_asm salg_avx512_asm \
-	salg_dble_avx512_asm pauli_avx512_asm pauli_dble_avx512_asm
-
 
 # Logging option (-mpilog or -mpitrace or -mpianim)
 
@@ -93,10 +88,10 @@ MDIR = ../modules
 VPATH = .:$(MDIR)/flags:$(MDIR)/lattice:$(MDIR)/archive:$(MDIR)/linalg:\
           $(MDIR)/random:$(MDIR)/uflds:$(MDIR)/mdflds:$(MDIR)/su3fcts:\
           $(MDIR)/utils:$(MDIR)/forces:$(MDIR)/sflds:$(MDIR)/dirac:\
-	  			$(MDIR)/sw_term:$(MDIR)/tcharge:$(MDIR)/block:$(MDIR)/sap:\
-	  			$(MDIR)/linsolv:$(MDIR)/dfl:$(MDIR)/vflds:$(MDIR)/little:\
+	  $(MDIR)/sw_term:$(MDIR)/tcharge:$(MDIR)/block:$(MDIR)/sap:\
+	  $(MDIR)/linsolv:$(MDIR)/dfl:$(MDIR)/vflds:$(MDIR)/little:\
           $(MDIR)/update:$(MDIR)/wflow:$(MDIR)/ratfcts:\
-					$(MDIR)/linalg/avx512:$(MDIR)/dirac/avx512:$(MDIR)/sw_term/avx512
+	  $(MDIR)/linalg/avx512:$(MDIR)/dirac/avx512:$(MDIR)/sw_term/avx512
 
 
 
@@ -116,7 +111,7 @@ LIBPATH = $(MPI_HOME)/lib
 
 CFLAGS = -std=c89 -pedantic -fstrict-aliasing \
          -Wall -Wno-long-long -Wstrict-prototypes \
-         -O2 -DAVX512
+         -O2 -DAVX512 -DAVX -DPM
 
 LFLAGS = $(CFLAGS)
 
@@ -128,7 +123,6 @@ LFLAGS = $(CFLAGS)
 #
 # -DCGNE_DBG -DFGCR_DBG -FGCR4VD_DBG -DMSCG_DBG
 # -DDFL_MODES_DBG -DMDINT_DBG -DRWRAT_DBG
-# -DAVX512 -DAVX512_ASM
 #
 # Add these (or some of these) to CFLAGS if debugging output is desired.
 
@@ -140,15 +134,7 @@ CLINKER=$(CC)
 
 
 #Check CFLAGS to find which AVX512 flags are active
-ifneq (,$(findstring AVX512_ASM,$(CFLAGS)))
-MODULES= $(STD_MODULES) $(AVX512_ASM_MODULES)
-else
-ifneq (,$(findstring AVX512,$(CFLAGS)))
-MODULES= $(STD_MODULES) $(AVX512_MODULES)
-else
 MODULES= $(STD_MODULES)
-endif
-endif
 
 PGMS= $(MAIN) $(MODULES)
 
@@ -159,18 +145,10 @@ PGMS= $(MAIN) $(MODULES)
 $(addsuffix .d,$(PGMS)): %.d: %.c Makefile
 	@ $(GCC) -ansi $< -MM $(addprefix -I,$(INCPATH)) -o $@
 
-
 # rule to compile source programs
 $(addsuffix .o,$(MAIN) $(STD_MODULES)): %.o: %.c Makefile
 	$(CC) $< -c $(CFLAGS) $(LOGOPTION) $(addprefix -I,$(INCPATH))
 
-# pattern to compile files in the avx512 directiories
-$(addsuffix .o,$(AVX512_MODULES)): %.o: %.c Makefile
-	$(CC) $< -c $(CFLAGS) $(LOGOPTION) $(addprefix -I,$(INCPATH))
-
-$(addsuffix .o,$(AVX512_ASM_MODULES)): %.o: %.s Makefile
-	$(CC) $< -c $(CFLAGS) $(LOGOPTION) $(addprefix -I,$(INCPATH))
-
 # rule to link object files
 $(MAIN): %: %.o $(addsuffix .o,$(MODULES)) Makefile
 	$(CLINKER) $< $(addsuffix .o,$(MODULES)) $(LFLAGS) $(LOGOPTION) \

modules/dirac/avx512/Dw_avx512.c

@@ -11,6 +11,8 @@
 * See ../Dw.c for more details and alternative implementations
  *******************************************************************************/
 
+#ifdef AVX512
+
 #include <stdlib.h>
 #include <stdio.h>
 #include <math.h>
@@ -215,3 +217,5 @@ void deo_avx512(int *piup, int *pidn, su3 *u, spinor *pl, float ceo, spin_t *rs)
   _avx512_dirac_combine_f_4( a3, b3 );
   _avx512_write_4_halfspinor_f_reverse_dn( a1, a2, a3, &(*sm).c3.c1.re, &(*sp).c3.c1.re, &(*sm2).c3.c1.re, &(*sp2).c3.c1.re );
 }
+
+#endif