1 /*
2
3 Copyright (C) 2014, The University of Texas at Austin
4
5 This file is part of libflame and is available under the 3-Clause
6 BSD license, which can be found in the LICENSE file at the top-level
7 directory, or at http://opensource.org/licenses/BSD-3-Clause
8
9 */
10 #include "FLAME.h"
11
12
13
14 #define N_VARIANTS 4
15
16 #define FLA_ALG_REFERENCE 0
17 #define FLA_ALG_UNBLOCKED 1
18 #define FLA_ALG_BLOCKED 2
19 #define FLA_ALG_OPTIMIZED 3
20
21
22 void time_Trmm_ruh(
23 int variant, int type, int n_repeats, int n, int nb_alg,
24 FLA_Obj A, FLA_Obj B, FLA_Obj C, FLA_Obj C_ref,
25 double *dtime, double *diff, double *gflops );
26
27
main(int argc,char * argv[])28 int main(int argc, char *argv[])
29 {
30 int
31 m_input, n_input,
32 m, n,
33 p_first, p_last, p_inc,
34 p,
35 nb_alg,
36 n_repeats,
37 variant,
38 i, j,
39 datatype,
40 n_variants = N_VARIANTS;
41
42 char *colors = "brkgmcbrkg";
43 char *ticks = "o+*xso+*xs";
44 char m_dim_desc[14];
45 char n_dim_desc[14];
46 char m_dim_tag[10];
47 char n_dim_tag[10];
48
49 double max_gflops=6.0;
50
51 double
52 dtime,
53 gflops,
54 diff;
55
56 FLA_Obj
57 A, B, C, C_ref;
58
59 /* Initialize FLAME */
60 FLA_Init( );
61
62
63 fprintf( stdout, "%c number of repeats: ", '%' );
64 scanf( "%d", &n_repeats );
65 fprintf( stdout, "%c %d\n", '%', n_repeats );
66
67 fprintf( stdout, "%c Enter blocking size: ", '%' );
68 scanf( "%d", &nb_alg );
69 fprintf( stdout, "%c %d\n", '%', nb_alg );
70
71 fprintf( stdout, "%c enter problem size first, last, inc: ", '%' );
72 scanf( "%d%d%d", &p_first, &p_last, &p_inc );
73 fprintf( stdout, "%c %d %d %d\n", '%', p_first, p_last, p_inc );
74
75 fprintf( stdout, "%c enter m n (-1 means bind to problem size): ", '%' );
76 scanf( "%d%d", &m_input, &n_input );
77 fprintf( stdout, "%c %d %d\n", '%', m_input, n_input );
78
79
80 /* Delete all existing data structures */
81 fprintf( stdout, "\nclear all;\n\n" );
82
83
84 if ( m_input > 0 ) {
85 sprintf( m_dim_desc, "m = %d", m_input );
86 sprintf( m_dim_tag, "m%dc", m_input);
87 }
88 else if( m_input < -1 ) {
89 sprintf( m_dim_desc, "m = p/%d", -m_input );
90 sprintf( m_dim_tag, "m%dp", -m_input );
91 }
92 else if( m_input == -1 ) {
93 sprintf( m_dim_desc, "m = p" );
94 sprintf( m_dim_tag, "m%dp", 1 );
95 }
96 if ( n_input > 0 ) {
97 sprintf( n_dim_desc, "n = %d", n_input );
98 sprintf( n_dim_tag, "n%dc", n_input);
99 }
100 else if( n_input < -1 ) {
101 sprintf( n_dim_desc, "n = p/%d", -n_input );
102 sprintf( n_dim_tag, "n%dp", -n_input );
103 }
104 else if( n_input == -1 ) {
105 sprintf( n_dim_desc, "n = p" );
106 sprintf( n_dim_tag, "n%dp", 1 );
107 }
108
109
110
111 for ( p = p_first, i = 1; p <= p_last; p += p_inc, i += 1 )
112 {
113
114 m = m_input;
115 n = n_input;
116
117 if( m < 0 ) m = p / f2c_abs(m_input);
118 if( n < 0 ) n = p / f2c_abs(n_input);
119
120 //datatype = FLA_COMPLEX;
121 datatype = FLA_DOUBLE_COMPLEX;
122
123 /* Allocate space for the matrices */
124 FLA_Obj_create( datatype, n, n, &A );
125 FLA_Obj_create( datatype, m, n, &C );
126 FLA_Obj_create( datatype, m, n, &C_ref );
127
128 /* Generate random matrices A, C */
129 FLA_Random_tri_matrix( FLA_UPPER_TRIANGULAR, FLA_UNIT_DIAG, A );
130 FLA_Random_matrix( C );
131
132 FLA_Copy_external( C, C_ref );
133
134
135 /* Time the reference implementation */
136 time_Trmm_ruh( 0, FLA_ALG_REFERENCE, n_repeats, p, nb_alg,
137 A, B, C, C_ref, &dtime, &diff, &gflops );
138
139 fprintf( stdout, "data_REF( %d, 1:2 ) = [ %d %6.3lf ]; \n", i, p, gflops );
140 fflush( stdout );
141
142 for ( variant = 1; variant <= n_variants; variant++ ){
143
144 //fprintf( stdout, "data_var%d( %d, 1:7 ) = [ %d ", variant, i, p );
145 fprintf( stdout, "data_var%d( %d, 1:5 ) = [ %d ", variant, i, p );
146 fflush( stdout );
147
148
149 time_Trmm_ruh( variant, FLA_ALG_UNBLOCKED, n_repeats, p, nb_alg,
150 A, B, C, C_ref, &dtime, &diff, &gflops );
151
152 fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
153 fflush( stdout );
154
155 time_Trmm_ruh( variant, FLA_ALG_BLOCKED, n_repeats, p, nb_alg,
156 A, B, C, C_ref, &dtime, &diff, &gflops );
157
158 fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
159 fflush( stdout );
160
161 //time_Trmm_ruh( variant, FLA_ALG_OPTIMIZED, n_repeats, p, nb_alg,
162 // A, B, C, C_ref, &dtime, &diff, &gflops );
163
164 //fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
165 //fflush( stdout );
166
167
168 fprintf( stdout, " ]; \n" );
169 fflush( stdout );
170 }
171 fprintf( stdout, "\n" );
172
173 FLA_Obj_free( &A );
174 FLA_Obj_free( &C );
175 FLA_Obj_free( &C_ref );
176 }
177
178 /*
179 fprintf( stdout, "figure;\n" );
180
181 fprintf( stdout, "plot( data_REF( :,1 ), data_REF( :, 2 ), '-' ); \n" );
182
183 fprintf( stdout, "hold on;\n" );
184
185 for ( i = 1; i <= n_variants; i++ ) {
186 fprintf( stdout, "plot( data_var%d( :,1 ), data_var%d( :, 2 ), '%c:%c' ); \n",
187 i, i, colors[ i-1 ], ticks[ i-1 ] );
188 fprintf( stdout, "plot( data_var%d( :,1 ), data_var%d( :, 4 ), '%c-.%c' ); \n",
189 i, i, colors[ i-1 ], ticks[ i-1 ] );
190 //fprintf( stdout, "plot( data_var%d( :,1 ), data_var%d( :, 6 ), '%c--%c' ); \n",
191 // i, i, colors[ i-1 ], ticks[ i-1 ] );
192 }
193
194 fprintf( stdout, "legend( ... \n" );
195 fprintf( stdout, "'Reference', ... \n" );
196
197 for ( i = 1; i < n_variants; i++ )
198 //fprintf( stdout, "'unb\\_var%d', 'blk\\_var%d', 'opt\\_var%d', ... \n", i, i, i );
199 fprintf( stdout, "'unb\\_var%d', 'blk\\_var%d', ... \n", i, i );
200
201 i = n_variants;
202 fprintf( stdout, "'unb\\_var%d', 'blk\\_var%d' ); \n", i, i );
203
204
205 fprintf( stdout, "xlabel( 'problem size p' );\n" );
206 fprintf( stdout, "ylabel( 'GFLOPS/sec.' );\n" );
207 fprintf( stdout, "axis( [ 0 %d 0 %.2f ] ); \n", p_last, max_gflops );
208 fprintf( stdout, "title( 'FLAME trmm\\_ruc performance (%s, %s)' );\n",
209 m_dim_desc, n_dim_desc );
210 fprintf( stdout, "print -depsc trmm_ruc_%s_%s.eps\n", m_dim_tag, n_dim_tag );
211 fprintf( stdout, "hold off;\n");
212 fflush( stdout );
213 */
214
215 FLA_Finalize( );
216 }
217
218