@@ -52,18 +52,20 @@ static void swap(int* m, int* n) { int tmp = *m; *m = *n; *n = tmp; }
5252
5353int main (int argc , char * argv [])
5454{
55- const int nrepeat = (1 < argc ? atoi (argv [1 ]) : 5 ), offset = 0 ;
55+ const int nrepeat = (1 < argc ? atoi (argv [1 ]) : 5 );
5656 const int nodd = (0 < nrepeat ? ((nrepeat & 1 /*odd*/ ) ? nrepeat : (nrepeat - 1 )) : 1 );
5757 const int stack_size = (2 < argc ? atoi (argv [2 ]) : 30000 );
5858 const int m = (3 < argc ? atoi (argv [3 ]) : 23 );
5959 const int n = (4 < argc ? atoi (argv [4 ]) : m );
60+ const int offset = (5 < argc ? atoi (argv [5 ]) : 0 );
61+ const int offset_stack_size = offset + stack_size ;
6062#if defined(ALIGNMENT ) && (0 < ALIGNMENT )
6163 const int mn = (int )ROUNDUP2 (sizeof (ELEM_TYPE ) * m , ALIGNMENT ) * n / sizeof (ELEM_TYPE );
6264#else
6365 const int mn = m * n ;
6466#endif
6567#if defined(SHUFFLE )
66- const size_t shuffle = libxsmm_shuffle ((unsigned int )stack_size );
68+ const size_t shuffle = libxsmm_shuffle ((unsigned int )offset_stack_size );
6769#endif
6870#if defined(WARMUP ) && (0 < WARMUP ) && !defined(_DEBUG )
6971 const int warmup = MAX (WARMUP , 2 ) / 2 * 2 ;
@@ -104,34 +106,34 @@ int main(int argc, char* argv[])
104106#else
105107 CHECK (acc_stream_create (& stream , "stream" , -1 /*default priority*/ ), & result );
106108#endif
107- CHECK (acc_host_mem_allocate ((void * * )& mat_hst , sizeof (ELEM_TYPE ) * mn * stack_size , stream ), & result );
108- CHECK (acc_host_mem_allocate ((void * * )& stack_hst , sizeof (int ) * stack_size , stream ), & result );
109+ CHECK (acc_host_mem_allocate ((void * * )& mat_hst , sizeof (ELEM_TYPE ) * mn * offset_stack_size , stream ), & result );
110+ CHECK (acc_host_mem_allocate ((void * * )& stack_hst , sizeof (int ) * offset_stack_size , stream ), & result );
109111 CHECK (acc_stream_sync (stream ), & result ); /* ensure host-data is allocated */
110- for (i = 0 ; i < stack_size ; ++ i ) { /* initialize matrices */
112+ for (i = 0 ; i < offset_stack_size ; ++ i ) { /* initialize matrices */
111113 init (i /*seed*/ , & mat_hst [i * mn ], m , n );
112114 }
113- for (i = 0 ; i < stack_size ; ++ i ) { /* initialize indexes */
115+ for (i = 0 ; i < offset_stack_size ; ++ i ) { /* initialize indexes */
114116#if defined(SHUFFLE )
115- const int j = mn * (int )((shuffle * i ) % stack_size );
117+ const int j = mn * (int )((shuffle * i ) % offset_stack_size );
116118#else
117119 const int j = mn * i ;
118120#endif
119121 stack_hst [i ] = j ;
120122 }
121- CHECK (acc_dev_mem_allocate ((void * * )& mat_dev , sizeof (ELEM_TYPE ) * mn * stack_size ), & result );
122- CHECK (acc_dev_mem_allocate ((void * * )& stack_dev , sizeof (int ) * stack_size ), & result );
123+ CHECK (acc_dev_mem_allocate ((void * * )& mat_dev , sizeof (ELEM_TYPE ) * mn * offset_stack_size ), & result );
124+ CHECK (acc_dev_mem_allocate ((void * * )& stack_dev , sizeof (int ) * offset_stack_size ), & result );
123125#if defined(USE_LIBXSMM )
124126 CHECK (acc_stream_sync (stream ), & result );
125127 start = libxsmm_timer_tick ();
126128#endif
127- CHECK (acc_memcpy_h2d (mat_hst , mat_dev , sizeof (ELEM_TYPE ) * mn * stack_size , stream ), & result );
128- CHECK (acc_memcpy_h2d (stack_hst , stack_dev , sizeof (int ) * stack_size , stream ), & result );
129+ CHECK (acc_memcpy_h2d (mat_hst , mat_dev , sizeof (ELEM_TYPE ) * mn * offset_stack_size , stream ), & result );
130+ CHECK (acc_memcpy_h2d (stack_hst , stack_dev , sizeof (int ) * offset_stack_size , stream ), & result );
129131#if defined(USE_LIBXSMM )
130132 CHECK (acc_stream_sync (stream ), & result );
131133 duration = libxsmm_timer_duration (start , libxsmm_timer_tick ());
132134 printf ("copy-in: %.1f ms %.1f GB/s\n" , 1000.0 * duration ,
133135 (sizeof (ELEM_TYPE ) * mn + sizeof (int ))
134- * stack_size / (duration * (1ULL << 30 )));
136+ * offset_stack_size / (duration * (1ULL << 30 )));
135137#endif
136138 /* warmup execution and prebuild JIT kernels */
137139 for (r = 0 ; r < warmup / 2 ; ++ r ) {
@@ -156,25 +158,25 @@ int main(int argc, char* argv[])
156158 assert (0 < nodd && (nodd & 1 /*odd*/ ));
157159 printf ("device: %.1f ms %.1f GB/s\n" , 1000.0 * duration / nodd ,
158160 (sizeof (ELEM_TYPE ) * mn + sizeof (int ))
159- * stack_size / (duration * (1ULL << 30 ) / nodd ));
161+ * offset_stack_size / (duration * (1ULL << 30 ) / nodd ));
160162 mm = m ; nn = n ;
161163 start = libxsmm_timer_tick ();
162164 for (r = 0 ; r < nodd ; ++ r ) {
163165 libxsmm_itrans_batch_omp (mat_hst , sizeof (ELEM_TYPE ), mm , nn , mm , nn ,
164- 0 /*index_base*/ , sizeof (int )/*index_stride*/ , stack_hst , stack_size );
166+ 0 /*index_base*/ , sizeof (int )/*index_stride*/ , stack_hst + offset , stack_size );
165167 swap (& mm , & nn );
166168 }
167169 duration = libxsmm_timer_duration (start , libxsmm_timer_tick ());
168170 printf ("host: %.1f ms %.1f GB/s\n" , 1000.0 * duration / nodd ,
169171 (sizeof (ELEM_TYPE ) * mn + sizeof (int ))
170- * stack_size / (duration * (1ULL << 30 ) / nodd ));
172+ * offset_stack_size / (duration * (1ULL << 30 ) / nodd ));
171173 /* transfer result from device to host for validation */
172174 CHECK (acc_memcpy_d2h (mat_dev , mat_hst ,
173- sizeof (ELEM_TYPE ) * mn * stack_size , stream ), & result );
175+ sizeof (ELEM_TYPE ) * mn * offset_stack_size , stream ), & result );
174176 CHECK (acc_stream_sync (stream ), & result );
175177 if (EXIT_SUCCESS == result ) {
176178 unsigned int nerrors = 0 ;
177- for (i = 0 ; i < stack_size ; ++ i ) {
179+ for (i = offset ; i < offset_stack_size ; ++ i ) {
178180 ELEM_TYPE gold [MAX_KERNEL_DIM * MAX_KERNEL_DIM ];
179181 const ELEM_TYPE * const test = mat_hst + mn * i ;
180182 init (i /*seed*/ , gold , m , n );
0 commit comments