root / BLAS / xGEMM / Makefile @ 1
Historique | Voir | Annoter | Télécharger (3,03 ko)
1 |
SOURCE=xGEMM.c |
---|---|
2 |
|
3 |
CC=gcc |
4 |
CFLAGS=-Wall -O3 |
5 |
LDFLAGS=-lm |
6 |
CUDADIR=/opt/cuda |
7 |
CUDASRC=$(CUDADIR)/src |
8 |
THUNKING=fortran_thunking.c |
9 |
CUDASRCINC=fortran_common.h |
10 |
CUDAINC=$(CUDADIR)/include |
11 |
CUDALIB=$(CUDADIR)/lib64 |
12 |
|
13 |
PATCHTHUNKING=patch_thunking.h |
14 |
|
15 |
GSLINC=/usr/include/gsl |
16 |
|
17 |
GOTO2=/opt/GotoBLAS2 |
18 |
|
19 |
ACML=/opt/acml |
20 |
ACMLINC=$(ACML)/gfortran64_mp/include |
21 |
ACMLLIB=$(ACML)/gfortran64_mp/lib |
22 |
|
23 |
|
24 |
EXECUTABLE=cblas fblas gsl cublas thunking gotoblas acml |
25 |
|
26 |
FORMAT=DOUBLE |
27 |
#FORMAT=FLOAT |
28 |
|
29 |
#DIRECTIVES=-D$(FORMAT) -DPRINT -DUNIT |
30 |
#DIRECTIVES=-D$(FORMAT) -DUNIT -DRESULTS -DQUIET |
31 |
DIRECTIVES=-DUNIT -DQUIET |
32 |
|
33 |
all: $(EXECUTABLE) |
34 |
|
35 |
cblas: $(SOURCE) |
36 |
|
37 |
$(CC) $(CFLAGS) $(DIRECTIVES) -DFLOAT -DCBLAS $(LDFLAGS) \ |
38 |
$(SOURCE) -lcblas -o $(SOURCE:.c=)_SP_$@ |
39 |
|
40 |
$(CC) $(CFLAGS) $(DIRECTIVES) -DDOUBLE -DCBLAS $(LDFLAGS) \ |
41 |
$(SOURCE) -lcblas -o $(SOURCE:.c=)_DP_$@ |
42 |
|
43 |
gotoblas: $(SOURCE) |
44 |
|
45 |
$(CC) $(CFLAGS) $(DIRECTIVES) -DFLOAT -DCBLAS $(LDFLAGS) \ |
46 |
$(SOURCE) $(GOTO2)/libgoto2.a -lpthread -o $(SOURCE:.c=)_SP_$@ |
47 |
|
48 |
$(CC) $(CFLAGS) $(DIRECTIVES) -DDOUBLE -DCBLAS $(LDFLAGS) \ |
49 |
$(SOURCE) $(GOTO2)/libgoto2.a -lpthread -o $(SOURCE:.c=)_DP_$@ |
50 |
|
51 |
acml: $(SOURCE) |
52 |
|
53 |
$(CC) -I$(ACMLINC) $(CFLAGS) $(DIRECTIVES) -DFLOAT -DACML $(LDFLAGS) \ |
54 |
$(SOURCE) -L$(ACMLLIB) -lacml_mp -lacml_mv \ |
55 |
-lgomp -lgfortran -lpthread -o $(SOURCE:.c=)_SP_$@ |
56 |
|
57 |
$(CC) -I$(ACMLINC) $(CFLAGS) $(DIRECTIVES) -DDOUBLE -DACML $(LDFLAGS) \ |
58 |
$(SOURCE) -L$(ACMLLIB) -lacml_mp -lacml_mv \ |
59 |
-lgomp -lgfortran -lpthread -o $(SOURCE:.c=)_DP_$@ |
60 |
|
61 |
fblas: $(SOURCE) |
62 |
|
63 |
$(CC) $(CFLAGS) $(DIRECTIVES) -DFLOAT -DFBLAS $(LDFLAGS) \ |
64 |
$(SOURCE) -lf77blas -o $(SOURCE:.c=)_SP_$@ |
65 |
|
66 |
$(CC) $(CFLAGS) $(DIRECTIVES) -DDOUBLE -DFBLAS $(LDFLAGS) \ |
67 |
$(SOURCE) -lf77blas -o $(SOURCE:.c=)_DP_$@ |
68 |
|
69 |
|
70 |
gsl: $(SOURCE) |
71 |
|
72 |
$(CC) -I$(GSLINC) $(CFLAGS) $(DIRECTIVES) -DFLOAT \ |
73 |
-DGSL $(LDFLAGS) \ |
74 |
$(SOURCE) -lgslcblas -o $(SOURCE:.c=)_SP_$@ |
75 |
|
76 |
$(CC) -I$(GSLINC) $(CFLAGS) $(DIRECTIVES) -DDOUBLE \ |
77 |
-DGSL $(LDFLAGS) \ |
78 |
$(SOURCE) -lgslcblas -o $(SOURCE:.c=)_DP_$@ |
79 |
|
80 |
cublas: $(SOURCE) |
81 |
|
82 |
$(CC) -I$(CUDAINC) -L$(CUDALIB) $(CFLAGS) -DFLOAT \ |
83 |
-DCUBLAS $(LDFLAGS) \ |
84 |
$(DIRECTIVES) $(SOURCE) -lcublas -o $(SOURCE:.c=)_SP_$@ |
85 |
|
86 |
$(CC) -I$(CUDAINC) -L$(CUDALIB) $(CFLAGS) -DDOUBLE \ |
87 |
-DCUBLAS $(LDFLAGS) \ |
88 |
$(DIRECTIVES) $(SOURCE) -lcublas -o $(SOURCE:.c=)_DP_$@ |
89 |
|
90 |
thunking: $(SOURCE) |
91 |
|
92 |
# Copy of source for thunking CUBLAS approach |
93 |
cp $(CUDASRC)/$(THUNKING) $(CUDASRC)/$(THUNKING:.c=.h) \ |
94 |
$(CUDASRC)/$(CUDASRCINC) . |
95 |
# Patch Thunking prototypes to compile on Debian Lenny |
96 |
patch $(CUDASRCINC) $(PATCHTHUNKING) |
97 |
$(CC) -I$(CUDAINC) $(CFLAGS) -c $(THUNKING) |
98 |
|
99 |
$(CC) -I$(CUDAINC) -L$(CUDALIB) $(CFLAGS) -DFLOAT -DTHUNKING \ |
100 |
$(LDFLAGS) $(DIRECTIVES) $(SOURCE) \ |
101 |
$(THUNKING:.c=.o) -lcublas -o $(SOURCE:.c=)_SP_$@ |
102 |
|
103 |
$(CC) -I$(CUDAINC) -L$(CUDALIB) $(CFLAGS) -DDOUBLE -DTHUNKING \ |
104 |
$(LDFLAGS) $(DIRECTIVES) $(SOURCE) \ |
105 |
$(THUNKING:.c=.o) -lcublas -o $(SOURCE:.c=)_DP_$@ |
106 |
|
107 |
clean: $(SOURCE) |
108 |
find . -name "$(SOURCE:.c=)_*" -exec rm {} \; |
109 |
find . -name "$(THUNKING:.c=)*" -exec rm {} \; |
110 |
find . -name "*~" -exec rm {} \; |
111 |
find . -name "$(CUDASRCINC)" -exec rm {} \; |
112 |
|
113 |
check: $(EXECUTABLE) |
114 |
|
115 |
$(SOURCE:.c=)_SP_$(EXECUTABLE) 2 1 |
116 |
|