root / BLAS / xGEMM / Makefile @ 53
Historique | Voir | Annoter | Télécharger (3,95 ko)
1 |
SOURCE=xGEMM.c |
---|---|
2 |
|
3 |
CC=gcc-4.7 |
4 |
FC=gfortran-4.7 |
5 |
CFLAGS=-Wall -O3 -ffast-math |
6 |
LDFLAGS=-lm |
7 |
CUDASRC=/usr/share/doc/nvidia-cuda-doc/examples/ |
8 |
THUNKING=fortran_thunking.c |
9 |
CUDASRCINC=fortran_common.h |
10 |
CUDAINC=/usr/include |
11 |
CUDALIB=/usr/lib/x86_64-linux-gnu/ |
12 |
# CLBLASINC=/opt/clBLAS/build/include |
13 |
# CLBLASLIB=/opt/clBLAS/build/library/ |
14 |
CLBLASINC=/opt/clBLAS/src |
15 |
CLBLASLIB=/opt/clBLAS/src/library/ |
16 |
|
17 |
PATCHTHUNKING=patch_thunking.h |
18 |
|
19 |
GSLINC=/usr/include/gsl |
20 |
|
21 |
GOTO2=/opt/GotoBLAS2 |
22 |
|
23 |
OPENBLAS=/usr/lib/openblas-base |
24 |
|
25 |
ACML=/opt/acml |
26 |
ACMLINC=$(ACML)/gfortran64_mp/include |
27 |
ACMLLIB=$(ACML)/gfortran64_mp/lib |
28 |
|
29 |
ATLAS=/usr/lib/atlas-base |
30 |
|
31 |
EXECUTABLE=cblas fblas gsl cublas thunking openblas acml clblas |
32 |
|
33 |
FORMAT=DOUBLE |
34 |
#FORMAT=FLOAT |
35 |
|
36 |
#DIRECTIVES=-D$(FORMAT) -DPRINT -DUNIT |
37 |
#DIRECTIVES=-D$(FORMAT) -DUNIT -DRESULTS -DQUIET |
38 |
DIRECTIVES=-DUNIT -DQUIET |
39 |
|
40 |
all: $(EXECUTABLE) |
41 |
|
42 |
cblas: $(SOURCE) |
43 |
|
44 |
$(CC) -L$(ATLAS) $(CFLAGS) $(DIRECTIVES) -DFLOAT -DCBLAS $(LDFLAGS) \ |
45 |
$(SOURCE) -latlas -lcblas -o $(SOURCE:.c=)_SP_$@ |
46 |
|
47 |
$(CC) -L$(ATLAS) $(CFLAGS) $(DIRECTIVES) -DDOUBLE -DCBLAS $(LDFLAGS) \ |
48 |
$(SOURCE) -latlas -lcblas -o $(SOURCE:.c=)_DP_$@ |
49 |
|
50 |
gotoblas: $(SOURCE) |
51 |
|
52 |
$(CC) $(CFLAGS) $(DIRECTIVES) -DFLOAT -DCBLAS $(LDFLAGS) \ |
53 |
$(SOURCE) $(GOTO2)/libgoto2.a -lpthread -o $(SOURCE:.c=)_SP_$@ |
54 |
|
55 |
$(CC) $(CFLAGS) $(DIRECTIVES) -DDOUBLE -DCBLAS $(LDFLAGS) \ |
56 |
$(SOURCE) $(GOTO2)/libgoto2.a -lpthread -o $(SOURCE:.c=)_DP_$@ |
57 |
|
58 |
openblas: $(SOURCE) |
59 |
|
60 |
$(CC) $(CFLAGS) $(DIRECTIVES) -fopenmp -DFLOAT -DCBLAS $(LDFLAGS) \ |
61 |
$(SOURCE) $(OPENBLAS)/libopenblas.a -lpthread -o $(SOURCE:.c=)_SP_$@ |
62 |
|
63 |
$(CC) $(CFLAGS) $(DIRECTIVES) -fopenmp -DDOUBLE -DCBLAS $(LDFLAGS) \ |
64 |
$(SOURCE) $(OPENBLAS)/libopenblas.a -lpthread -o $(SOURCE:.c=)_DP_$@ |
65 |
|
66 |
acml: $(SOURCE) |
67 |
|
68 |
$(CC) -I$(ACMLINC) $(CFLAGS) $(DIRECTIVES) -DFLOAT -DACML $(LDFLAGS) \ |
69 |
$(SOURCE) -L$(ACMLLIB) -lacml_mp \ |
70 |
-lgomp -lgfortran -lpthread -o $(SOURCE:.c=)_SP_$@ |
71 |
|
72 |
$(CC) -I$(ACMLINC) $(CFLAGS) $(DIRECTIVES) -DDOUBLE -DACML $(LDFLAGS) \ |
73 |
$(SOURCE) -L$(ACMLLIB) -lacml_mp \ |
74 |
-lgomp -lgfortran -lpthread -o $(SOURCE:.c=)_DP_$@ |
75 |
|
76 |
fblas: $(SOURCE) |
77 |
|
78 |
$(CC) -L$(ATLAS) $(CFLAGS) $(DIRECTIVES) -DFLOAT -DFBLAS $(LDFLAGS) \ |
79 |
$(SOURCE) -latlas -lf77blas -o $(SOURCE:.c=)_SP_$@ |
80 |
|
81 |
$(CC) -L$(ATLAS) $(CFLAGS) $(DIRECTIVES) -DDOUBLE -DFBLAS $(LDFLAGS) \ |
82 |
$(SOURCE) -latlas -lf77blas -o $(SOURCE:.c=)_DP_$@ |
83 |
|
84 |
|
85 |
gsl: $(SOURCE) |
86 |
|
87 |
$(CC) -I$(GSLINC) $(CFLAGS) $(DIRECTIVES) -DFLOAT \ |
88 |
-DGSL $(LDFLAGS) \ |
89 |
$(SOURCE) -lgslcblas -o $(SOURCE:.c=)_SP_$@ |
90 |
|
91 |
$(CC) -I$(GSLINC) $(CFLAGS) $(DIRECTIVES) -DDOUBLE \ |
92 |
-DGSL $(LDFLAGS) \ |
93 |
$(SOURCE) -lgslcblas -o $(SOURCE:.c=)_DP_$@ |
94 |
|
95 |
clblas: $(SOURCE) |
96 |
|
97 |
$(CC) -I$(CLBLASINC) -L$(CLBLASLIB) $(CFLAGS) -DFLOAT \ |
98 |
-DCLBLAS $(LDFLAGS) \ |
99 |
$(DIRECTIVES) $(SOURCE) -lclBLAS -lOpenCL -o $(SOURCE:.c=)_SP_$@ |
100 |
|
101 |
$(CC) -I$(CLBLASINC) -L$(CLBLASLIB) $(CFLAGS) -DDOUBLE \ |
102 |
-DCLBLAS $(LDFLAGS) \ |
103 |
$(DIRECTIVES) $(SOURCE) -lclBLAS -lOpenCL -o $(SOURCE:.c=)_DP_$@ |
104 |
|
105 |
cublas: $(SOURCE) |
106 |
|
107 |
$(CC) -I$(CUDAINC) -L$(CUDALIB) $(CFLAGS) -DFLOAT \ |
108 |
-DCUBLAS $(LDFLAGS) \ |
109 |
$(DIRECTIVES) $(SOURCE) -lcublas -o $(SOURCE:.c=)_SP_$@ |
110 |
|
111 |
$(CC) -I$(CUDAINC) -L$(CUDALIB) $(CFLAGS) -DDOUBLE \ |
112 |
-DCUBLAS $(LDFLAGS) \ |
113 |
$(DIRECTIVES) $(SOURCE) -lcublas -o $(SOURCE:.c=)_DP_$@ |
114 |
|
115 |
thunking: $(SOURCE) |
116 |
|
117 |
# Copy of source for thunking CUBLAS approach |
118 |
cp $(CUDASRC)/$(THUNKING) $(CUDASRC)/$(THUNKING:.c=.h) \ |
119 |
$(CUDASRC)/$(CUDASRCINC) . |
120 |
# Patch Thunking prototypes to compile on Debian Lenny |
121 |
$(CC) -DCUBLAS_GFORTRAN -I$(CUDAINC) $(CFLAGS) -c $(THUNKING) |
122 |
|
123 |
$(CC) -DCUBLAS_GFORTRAN \ |
124 |
-I$(CUDAINC) -L$(CUDALIB) $(CFLAGS) -DFLOAT -DTHUNKING \ |
125 |
$(LDFLAGS) $(DIRECTIVES) $(SOURCE) \ |
126 |
$(THUNKING:.c=.o) -lcublas -o $(SOURCE:.c=)_SP_$@ |
127 |
|
128 |
$(CC) -DCUBLAS_GFORTRAN \ |
129 |
-I$(CUDAINC) -L$(CUDALIB) $(CFLAGS) -DDOUBLE -DTHUNKING \ |
130 |
$(LDFLAGS) $(DIRECTIVES) $(SOURCE) \ |
131 |
$(THUNKING:.c=.o) -lcublas -o $(SOURCE:.c=)_DP_$@ |
132 |
|
133 |
clean: $(SOURCE) |
134 |
find . -name "$(SOURCE:.c=)_*" -exec rm {} \; |
135 |
find . -name "$(THUNKING:.c=)*" -exec rm {} \; |
136 |
find . -name "*~" -exec rm {} \; |
137 |
find . -name "$(CUDASRCINC)" -exec rm {} \; |
138 |
|
139 |
check: $(EXECUTABLE) |
140 |
|
141 |
$(SOURCE:.c=)_SP_$(EXECUTABLE) 2 1 |
142 |
|