root / BLAS / xGEMM / Makefile @ 146
Historique | Voir | Annoter | Télécharger (4,18 ko)
1 |
SOURCE=xGEMM.c |
---|---|
2 |
|
3 |
CC=gcc |
4 |
FC=gfortran |
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=/usr/include |
13 |
CLBLASLIB=/usr/lib |
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 |
OPENBLAS=/usr/lib |
25 |
|
26 |
ACML=/opt/acml |
27 |
ACMLINC=$(ACML)/gfortran64_mp/include |
28 |
ACMLLIB=$(ACML)/gfortran64_mp/lib |
29 |
|
30 |
ATLAS=/usr/lib/atlas-base |
31 |
|
32 |
#EXECUTABLE=cblas fblas gsl cublas thunking openblas acml clblas |
33 |
#EXECUTABLE=cblas fblas gsl cublas thunking openblas clblas |
34 |
#EXECUTABLE=gsl openblas clblas cublas thunking |
35 |
#EXECUTABLE=openblas clblas cublas thunking |
36 |
#EXECUTABLE=clblas cublas thunking |
37 |
EXECUTABLE=openblas |
38 |
|
39 |
FORMAT=DOUBLE |
40 |
#FORMAT=FLOAT |
41 |
|
42 |
#DIRECTIVES=-D$(FORMAT) -DPRINT -DUNIT |
43 |
#DIRECTIVES=-D$(FORMAT) -DUNIT -DRESULTS -DQUIET |
44 |
DIRECTIVES=-DUNIT -DQUIET |
45 |
|
46 |
all: $(EXECUTABLE) |
47 |
|
48 |
cblas: $(SOURCE) |
49 |
|
50 |
$(CC) -L$(ATLAS) $(CFLAGS) $(DIRECTIVES) -DFLOAT -DCBLAS $(LDFLAGS) \ |
51 |
$(SOURCE) -latlas -lcblas -o $(SOURCE:.c=)_SP_$@ |
52 |
|
53 |
$(CC) -L$(ATLAS) $(CFLAGS) $(DIRECTIVES) -DDOUBLE -DCBLAS $(LDFLAGS) \ |
54 |
$(SOURCE) -latlas -lcblas -o $(SOURCE:.c=)_DP_$@ |
55 |
|
56 |
gotoblas: $(SOURCE) |
57 |
|
58 |
$(CC) $(CFLAGS) $(DIRECTIVES) -DFLOAT -DCBLAS $(LDFLAGS) \ |
59 |
$(SOURCE) $(GOTO2)/libgoto2.a -lpthread -o $(SOURCE:.c=)_SP_$@ |
60 |
|
61 |
$(CC) $(CFLAGS) $(DIRECTIVES) -DDOUBLE -DCBLAS $(LDFLAGS) \ |
62 |
$(SOURCE) $(GOTO2)/libgoto2.a -lpthread -o $(SOURCE:.c=)_DP_$@ |
63 |
|
64 |
openblas: $(SOURCE) |
65 |
|
66 |
$(CC) $(CFLAGS) $(DIRECTIVES) -L/usr/include/openblas -fopenmp -DFLOAT -DCBLAS $(LDFLAGS) \ |
67 |
$(SOURCE) $(OPENBLAS)/libopenblas.a -lpthread -o $(SOURCE:.c=)_SP_$@ |
68 |
|
69 |
$(CC) $(CFLAGS) $(DIRECTIVES) -L/usr/include/openblas -fopenmp -DDOUBLE -DCBLAS $(LDFLAGS) \ |
70 |
$(SOURCE) $(OPENBLAS)/libopenblas.a -lpthread -o $(SOURCE:.c=)_DP_$@ |
71 |
|
72 |
acml: $(SOURCE) |
73 |
|
74 |
$(CC) -I$(ACMLINC) $(CFLAGS) $(DIRECTIVES) -DFLOAT -DACML $(LDFLAGS) \ |
75 |
$(SOURCE) -L$(ACMLLIB) -lacml_mp \ |
76 |
-lgomp -lgfortran -lpthread -o $(SOURCE:.c=)_SP_$@ |
77 |
|
78 |
$(CC) -I$(ACMLINC) $(CFLAGS) $(DIRECTIVES) -DDOUBLE -DACML $(LDFLAGS) \ |
79 |
$(SOURCE) -L$(ACMLLIB) -lacml_mp \ |
80 |
-lgomp -lgfortran -lpthread -o $(SOURCE:.c=)_DP_$@ |
81 |
|
82 |
fblas: $(SOURCE) |
83 |
|
84 |
$(CC) -L$(ATLAS) $(CFLAGS) $(DIRECTIVES) -DFLOAT -DFBLAS $(LDFLAGS) \ |
85 |
$(SOURCE) -latlas -lf77blas -o $(SOURCE:.c=)_SP_$@ |
86 |
|
87 |
$(CC) -L$(ATLAS) $(CFLAGS) $(DIRECTIVES) -DDOUBLE -DFBLAS $(LDFLAGS) \ |
88 |
$(SOURCE) -latlas -lf77blas -o $(SOURCE:.c=)_DP_$@ |
89 |
|
90 |
|
91 |
gsl: $(SOURCE) |
92 |
|
93 |
$(CC) -I$(GSLINC) $(CFLAGS) $(DIRECTIVES) -DFLOAT \ |
94 |
-DGSL $(LDFLAGS) \ |
95 |
$(SOURCE) -lgslcblas -o $(SOURCE:.c=)_SP_$@ |
96 |
|
97 |
$(CC) -I$(GSLINC) $(CFLAGS) $(DIRECTIVES) -DDOUBLE \ |
98 |
-DGSL $(LDFLAGS) \ |
99 |
$(SOURCE) -lgslcblas -o $(SOURCE:.c=)_DP_$@ |
100 |
|
101 |
clblas: $(SOURCE) |
102 |
|
103 |
$(CC) -I$(CLBLASINC) -L$(CLBLASLIB) $(CFLAGS) -DFLOAT \ |
104 |
-DCLBLAS $(LDFLAGS) \ |
105 |
$(DIRECTIVES) $(SOURCE) -lclBLAS -lOpenCL -o $(SOURCE:.c=)_SP_$@ |
106 |
|
107 |
$(CC) -I$(CLBLASINC) -L$(CLBLASLIB) $(CFLAGS) -DDOUBLE \ |
108 |
-DCLBLAS $(LDFLAGS) \ |
109 |
$(DIRECTIVES) $(SOURCE) -lclBLAS -lOpenCL -o $(SOURCE:.c=)_DP_$@ |
110 |
|
111 |
cublas: $(SOURCE) |
112 |
|
113 |
$(CC) -I$(CUDAINC) -L$(CUDALIB) $(CFLAGS) -DFLOAT \ |
114 |
-DCUBLAS $(LDFLAGS) \ |
115 |
$(DIRECTIVES) $(SOURCE) -lcublas -o $(SOURCE:.c=)_SP_$@ |
116 |
|
117 |
$(CC) -I$(CUDAINC) -L$(CUDALIB) $(CFLAGS) -DDOUBLE \ |
118 |
-DCUBLAS $(LDFLAGS) \ |
119 |
$(DIRECTIVES) $(SOURCE) -lcublas -o $(SOURCE:.c=)_DP_$@ |
120 |
|
121 |
thunking: $(SOURCE) |
122 |
|
123 |
# Copy of source for thunking CUBLAS approach |
124 |
cp $(CUDASRC)/$(THUNKING) $(CUDASRC)/$(THUNKING:.c=.h) \ |
125 |
$(CUDASRC)/$(CUDASRCINC) . |
126 |
# Patch Thunking prototypes to compile on Debian Lenny |
127 |
$(CC) -DCUBLAS_GFORTRAN -I$(CUDAINC) $(CFLAGS) -c $(THUNKING) |
128 |
|
129 |
$(CC) -DCUBLAS_GFORTRAN \ |
130 |
-I$(CUDAINC) -L$(CUDALIB) $(CFLAGS) -DFLOAT -DTHUNKING \ |
131 |
$(LDFLAGS) $(DIRECTIVES) $(SOURCE) \ |
132 |
$(THUNKING:.c=.o) -lcublas -o $(SOURCE:.c=)_SP_$@ |
133 |
|
134 |
$(CC) -DCUBLAS_GFORTRAN \ |
135 |
-I$(CUDAINC) -L$(CUDALIB) $(CFLAGS) -DDOUBLE -DTHUNKING \ |
136 |
$(LDFLAGS) $(DIRECTIVES) $(SOURCE) \ |
137 |
$(THUNKING:.c=.o) -lcublas -o $(SOURCE:.c=)_DP_$@ |
138 |
|
139 |
clean: $(SOURCE) |
140 |
find . -name "$(SOURCE:.c=)_*" -exec rm {} \; |
141 |
find . -name "$(THUNKING:.c=)*" -exec rm {} \; |
142 |
find . -name "*~" -exec rm {} \; |
143 |
find . -name "$(CUDASRCINC)" -exec rm {} \; |
144 |
|
145 |
check: $(EXECUTABLE) |
146 |
|
147 |
$(SOURCE:.c=)_SP_$(EXECUTABLE) 2 1 |
148 |
|