Context Navigation

source: soft/giet_vm/applications/transpose/transpose.c @ 777

Last change on this file since 777 was 775, checked in by alain, 9 years ago
Introduce the dump command in shell.
File size: 21.1 KB

Line
1	//////////////////////////////////////////////////////////////////////////////////////////
2	// File : transpose.c
3	// Date : september 2015
4	// author : Alain Greiner
5	//////////////////////////////////////////////////////////////////////////////////////////
6	// This multi-threaded aplication read a raw image (one byte per pixel)
7	// stored on disk, transpose it, display the result on the frame buffer,
8	// and store the transposed image on disk.
9	// The input image can be interactively selected if the INTERACTIVE flag is set.
10	// It can run on a multi-processors, multi-clusters architecture, with one thread
11	// per processor, and uses the POSIX threads API.
12	// It uses the giet_fat_mmap() to directly access the input and output files
13	// in the kernel files cache. It does not use the CMA to display the result image.
14	//
15	// The main() function can be launched on any processor P[x,y,l].
16	// It makes the initialisations, launch (N-1) threads to run the execute() function
17	// on the (N-1) other processors than P[x,y,l], call himself the execute() function,
18	// and finally call the instrument() function to display instrumentation results
19	// when the parallel execution is completed.
20	//
21	// The buf_in[x,y] and buf_out[put buffers containing the direct ans transposed images
22	// are distributed in clusters:
23	// In each cluster[x,y], the thread running on processor P[x,y,0] uses the giet_fat_mmap()
24	// function to map the buf_in[x,y] and buf_out[x,y] buffers containing a set of lines.
25	// Then, all threads in cluster[x,y] read pixels from the local buf_in[x,y] buffer, and
26	// write the pixels to the remote buf_out[x,y] buffers. Finally, each thread display
27	// a part of the transposed image to the frame buffer.
28	//
29	// - The image size must fit the frame buffer width and height, that must be power of 2.
30	// - The number of clusters must be a power of 2 no larger than 256.
31	// - The number of processors per cluster must be a power of 2 no larger than 4.
32	// - The number of clusters cannot be larger than (image_size * image_size) / 4096,
33	// because the size of buf_in[x,y] and buf_out[x,y] must be multiple of 4096.
34	//
35	// The transpose_rw.c file contains a variant that use the giet_fat_read()
36	// and giet_fat_write() system calls, to access the files.
37	//////////////////////////////////////////////////////////////////////////////////////////
38
39	#include "stdio.h"
40	#include "stdlib.h"
41	#include "user_barrier.h"
42	#include "malloc.h"
43
44	#define BLOCK_SIZE 512 // block size on disk
45	#define X_MAX 16 // max number of clusters in row
46	#define Y_MAX 16 // max number of clusters in column
47	#define PROCS_MAX 4 // max number of procs per cluster
48	#define CLUSTER_MAX (X_MAX * Y_MAX) // max number of clusters
49	#define IMAGE_SIZE 128 // default image size
50	#define INPUT_FILE_PATH "/misc/images_128.raw" // default input file pathname
51	#define OUTPUT_FILE_PATH "/home/trsp_128.raw" // default output file pathname
52	#define INTERACTIVE 0 // interactive capture of filenames
53	#define VERBOSE 0 // print comments on TTY
54
55
56	// macro to use a shared TTY
57	#define printf(...); { lock_acquire( &tty_lock ); \
58	giet_tty_printf(__VA_ARGS__); \
59	lock_release( &tty_lock ); }
60
61	///////////////////////////////////////////////////////
62	// global variables stored in seg_data in cluster(0,0)
63	///////////////////////////////////////////////////////
64
65	// instrumentation counters for each processor in each cluster
66	unsigned int MMAP_START[X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
67	unsigned int MMAP_END [X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
68	unsigned int TRSP_START[X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
69	unsigned int TRSP_END [X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
70	unsigned int DISP_START[X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
71	unsigned int DISP_END [X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
72
73	// arrays of pointers on distributed buffers
74	// one input buffer & one output buffer per cluster
75	unsigned char* buf_in [CLUSTER_MAX];
76	unsigned char* buf_out[CLUSTER_MAX];
77
78	// lock protecting shared TTY
79	user_lock_t tty_lock;
80
81	// synchronisation barrier (all threads)
82	giet_sqt_barrier_t barrier;
83
84	// input & output files pathname and size
85	char input_file_name[256];
86	char output_file_name[256];
87	unsigned int image_size;
88
89	// input & output file descriptors
90	int fd_in;
91	int fd_out;
92
93	////////////////////////////////////////////
94	__attribute__ ((constructor)) void execute()
95	////////////////////////////////////////////
96	{
97	unsigned int l; // line index for loops
98	unsigned int p; // pixel index for loops
99
100	// get processor identifiers
101	unsigned int x_id; // x cluster coordinate
102	unsigned int y_id; // y cluster coordinate
103	unsigned int p_id; // local processor index
104
105	giet_proc_xyp( &x_id, &y_id, &p_id);
106
107	// get & check plat-form parameters
108	unsigned int x_size; // number of clusters in a row
109	unsigned int y_size; // number of clusters in a column
110	unsigned int nprocs; // number of processors per cluster
111
112	giet_procs_number( &x_size , &y_size , &nprocs );
113
114	unsigned int nclusters = x_size * y_size; // number of clusters
115	unsigned int nthreads = x_size * y_size * nprocs; // number of threads
116	unsigned int npixels = image_size * image_size; // pixels per image
117	unsigned int cluster_id = (x_id * y_size) + y_id; // "continuous" index
118	unsigned int thread_id = (cluster_id * nprocs) + p_id; // "continuous" index
119
120	// parallel load of image:
121	// thread running on processor[x,y,0]
122	// map input & output files in buf_in & buf_out buffers.
123
124	MMAP_START[x_id][y_id][p_id] = giet_proctime();
125
126	if ( p_id == 0 )
127	{
128	// map buf_in
129	unsigned int length = npixels / nclusters;
130	unsigned int offset = length * cluster_id;
131
132	buf_in[cluster_id] = giet_fat_mmap( NULL,
133	length,
134	MAP_PROT_READ,
135	MAP_SHARED,
136	fd_in,
137	offset );
138	if ( buf_in[cluster_id] == NULL )
139	{
140	printf("\n[TRANSPOSE ERROR] Thread[%d,%d,%d] cannot map input file\n",
141	x_id , y_id , p_id );
142	giet_pthread_exit( NULL );
143	}
144
145	if ( VERBOSE )
146	{
147	printf("\n@@@ Thread[%d,%d,%d] call mmap for input file\n"
148	" length = %x / offset = %x / buf_in = %x\n",
149	x_id , y_id , p_id , length , offset , buf_in[cluster_id] );
150	}
151
152	// map buf_out
153	buf_out[cluster_id] = giet_fat_mmap( NULL,
154	length,
155	MAP_PROT_WRITE,
156	MAP_SHARED,
157	fd_out,
158	offset );
159
160	if ( buf_out[cluster_id] == NULL )
161	{
162	printf("\n[TRANSPOSE ERROR] Thread[%d,%d,%d] cannot map output file\n",
163	x_id , y_id , p_id );
164	giet_pthread_exit( NULL );
165	}
166
167	if ( VERBOSE )
168	{
169	printf("\n@@@ Thread[%d,%d,%d] call mmap for output file\n"
170	" length = %x / offset = %x / buf_out = %x\n",
171	x_id , y_id , p_id , length , offset , buf_out[cluster_id] );
172	}
173
174	}
175
176	MMAP_END[x_id][y_id][p_id] = giet_proctime();
177
178	/////////////////////////////
179	sqt_barrier_wait( &barrier );
180	/////////////////////////////
181
182	// parallel transpose from buf_in to buf_out
183	// each thread makes the transposition for nlt lines (nlt = image_size/nthreads)
184	// from line [thread_idnlt] to line [(thread_id + 1)nlt - 1]
185	// (p,l) are the absolute pixel coordinates in the source image
186
187	TRSP_START[x_id][y_id][p_id] = giet_proctime();
188
189	unsigned int nlt = image_size / nthreads; // number of lines per thread
190	unsigned int nlc = image_size / nclusters; // number of lines per cluster
191
192	unsigned int src_cluster;
193	unsigned int src_index;
194	unsigned int dst_cluster;
195	unsigned int dst_index;
196
197	unsigned char byte;
198
199	unsigned int first = thread_id * nlt; // first line index for a given thread
200	unsigned int last = first + nlt; // last line index for a given thread
201
202	for ( l = first ; l < last ; l++ )
203	{
204	// in each iteration we transfer one byte
205	for ( p = 0 ; p < image_size ; p++ )
206	{
207	// read one byte from local buf_in
208	src_cluster = l / nlc;
209	src_index = (l % nlc)*image_size + p;
210	byte = buf_in[src_cluster][src_index];
211
212	// write one byte to remote buf_out
213	dst_cluster = p / nlc;
214	dst_index = (p % nlc)*image_size + l;
215	buf_out[dst_cluster][dst_index] = byte;
216	}
217	}
218
219	if ( (p_id == 0) && (x_id==0) && (y_id==0) )
220	{
221	printf("\n[TRANSPOSE] Thread[%d,%d,%d] completes transpose at cycle %d\n",
222	x_id, y_id, p_id, giet_proctime() );
223	}
224
225	TRSP_END[x_id][y_id][p_id] = giet_proctime();
226
227	/////////////////////////////
228	sqt_barrier_wait( &barrier );
229	/////////////////////////////
230
231	// parallel display from local buf_out to frame buffer
232	// all threads contribute to display using memcpy...
233
234	DISP_START[x_id][y_id][p_id] = giet_proctime();
235
236	unsigned int npt = npixels / nthreads; // number of pixels per thread
237
238	giet_fbf_sync_write( npt * thread_id,
239	&buf_out[cluster_id][p_id*npt],
240	npt );
241
242	if ( (x_id==0) && (y_id==0) && (p_id==0) )
243	{
244	printf("\n[TRANSPOSE] Thread[%d,%d,%d] completes display at cycle %d\n",
245	x_id, y_id, p_id, giet_proctime() );
246	}
247
248	DISP_END[x_id][y_id][p_id] = giet_proctime();
249
250	/////////////////////////////
251	sqt_barrier_wait( &barrier );
252	/////////////////////////////
253
254	// all threads, but thread[0,0,0], suicide
255	if ( (x_id != 0) \|\| (y_id != 0) \|\| (p_id != 0) )
256	giet_pthread_exit( "completed" );
257
258	} // end execute()
259
260
261
262	//////////////////////////////////////
263	void instrument( unsigned int x_size,
264	unsigned int y_size,
265	unsigned int nprocs )
266	//////////////////////////////////////
267	{
268	unsigned int x, y, l;
269
270	unsigned int min_load_start = 0xFFFFFFFF;
271	unsigned int max_load_start = 0;
272	unsigned int min_load_ended = 0xFFFFFFFF;
273	unsigned int max_load_ended = 0;
274	unsigned int min_trsp_start = 0xFFFFFFFF;
275	unsigned int max_trsp_start = 0;
276	unsigned int min_trsp_ended = 0xFFFFFFFF;
277	unsigned int max_trsp_ended = 0;
278	unsigned int min_disp_start = 0xFFFFFFFF;
279	unsigned int max_disp_start = 0;
280	unsigned int min_disp_ended = 0xFFFFFFFF;
281	unsigned int max_disp_ended = 0;
282
283	char string[64];
284
285	snprintf( string , 64 , "/home/transpose_%d_%d_%d" , x_size , y_size , nprocs );
286
287	// open instrumentation file
288	unsigned int fd = giet_fat_open( string , O_CREATE );
289	if ( fd < 0 )
290	{
291	printf("\n[TRANSPOSE ERROR] cannot open instrumentation file %s\n", string );
292	giet_pthread_exit( NULL );
293	}
294
295	for (x = 0; x < x_size; x++)
296	{
297	for (y = 0; y < y_size; y++)
298	{
299	for ( l = 0 ; l < nprocs ; l++ )
300	{
301	if (MMAP_START[x][y][l] < min_load_start) min_load_start = MMAP_START[x][y][l];
302	if (MMAP_START[x][y][l] > max_load_start) max_load_start = MMAP_START[x][y][l];
303	if (MMAP_END[x][y][l] < min_load_ended) min_load_ended = MMAP_END[x][y][l];
304	if (MMAP_END[x][y][l] > max_load_ended) max_load_ended = MMAP_END[x][y][l];
305	if (TRSP_START[x][y][l] < min_trsp_start) min_trsp_start = TRSP_START[x][y][l];
306	if (TRSP_START[x][y][l] > max_trsp_start) max_trsp_start = TRSP_START[x][y][l];
307	if (TRSP_END[x][y][l] < min_trsp_ended) min_trsp_ended = TRSP_END[x][y][l];
308	if (TRSP_END[x][y][l] > max_trsp_ended) max_trsp_ended = TRSP_END[x][y][l];
309	if (DISP_START[x][y][l] < min_disp_start) min_disp_start = DISP_START[x][y][l];
310	if (DISP_START[x][y][l] > max_disp_start) max_disp_start = DISP_START[x][y][l];
311	if (DISP_END[x][y][l] < min_disp_ended) min_disp_ended = DISP_END[x][y][l];
312	if (DISP_END[x][y][l] > max_disp_ended) max_disp_ended = DISP_END[x][y][l];
313	}
314	}
315	}
316
317	giet_tty_printf( "\n ------ %s ------\n" , string );
318	giet_fat_fprintf( fd , "\n ------ %s ------\n" , string );
319
320	giet_tty_printf( " - MMAP_START : min = %d / max = %d / med = %d / delta = %d\n",
321	min_load_start, max_load_start, (min_load_start+max_load_start)/2,
322	max_load_start-min_load_start );
323
324	giet_fat_fprintf( fd , " - MMAP_START : min = %d / max = %d / med = %d / delta = %d\n",
325	min_load_start, max_load_start, (min_load_start+max_load_start)/2,
326	max_load_start-min_load_start );
327
328	giet_tty_printf( " - MMAP_END : min = %d / max = %d / med = %d / delta = %d\n",
329	min_load_ended, max_load_ended, (min_load_ended+max_load_ended)/2,
330	max_load_ended-min_load_ended );
331
332	giet_fat_fprintf( fd , " - MMAP_END : min = %d / max = %d / med = %d / delta = %d\n",
333	min_load_ended, max_load_ended, (min_load_ended+max_load_ended)/2,
334	max_load_ended-min_load_ended );
335
336	giet_tty_printf( " - TRSP_START : min = %d / max = %d / med = %d / delta = %d\n",
337	min_trsp_start, max_trsp_start, (min_trsp_start+max_trsp_start)/2,
338	max_trsp_start-min_trsp_start );
339
340	giet_fat_fprintf( fd , " - TRSP_START : min = %d / max = %d / med = %d / delta = %d\n",
341	min_trsp_start, max_trsp_start, (min_trsp_start+max_trsp_start)/2,
342	max_trsp_start-min_trsp_start );
343
344	giet_tty_printf( " - TRSP_END : min = %d / max = %d / med = %d / delta = %d\n",
345	min_trsp_ended, max_trsp_ended, (min_trsp_ended+max_trsp_ended)/2,
346	max_trsp_ended-min_trsp_ended );
347
348	giet_fat_fprintf( fd , " - TRSP_END : min = %d / max = %d / med = %d / delta = %d\n",
349	min_trsp_ended, max_trsp_ended, (min_trsp_ended+max_trsp_ended)/2,
350	max_trsp_ended-min_trsp_ended );
351
352	giet_tty_printf( " - DISP_START : min = %d / max = %d / med = %d / delta = %d\n",
353	min_disp_start, max_disp_start, (min_disp_start+max_disp_start)/2,
354	max_disp_start-min_disp_start );
355
356	giet_fat_fprintf( fd , " - DISP_START : min = %d / max = %d / med = %d / delta = %d\n",
357	min_disp_start, max_disp_start, (min_disp_start+max_disp_start)/2,
358	max_disp_start-min_disp_start );
359
360	giet_tty_printf( " - DISP_END : min = %d / max = %d / med = %d / delta = %d\n",
361	min_disp_ended, max_disp_ended, (min_disp_ended+max_disp_ended)/2,
362	max_disp_ended-min_disp_ended );
363
364	giet_fat_fprintf( fd , " - DISP_END : min = %d / max = %d / med = %d / delta = %d\n",
365	min_disp_ended, max_disp_ended, (min_disp_ended+max_disp_ended)/2,
366	max_disp_ended-min_disp_ended );
367
368	giet_fat_close( fd );
369
370	} // end instrument()
371
372
373
374	//////////////////////////////////////////
375	__attribute__ ((constructor)) void main()
376	//////////////////////////////////////////
377	{
378	// indexes for loops
379	unsigned int x , y , n;
380
381	// get identifiers for proc executing main
382	unsigned int x_id; // x cluster coordinate
383	unsigned int y_id; // y cluster coordinate
384	unsigned int p_id; // local processor index
385	giet_proc_xyp( &x_id , &y_id , &p_id );
386
387	// get & check plat-form parameters
388	unsigned int x_size; // number of clusters in a row
389	unsigned int y_size; // number of clusters in a column
390	unsigned int nprocs; // number of processors per cluster
391	giet_procs_number( &x_size , &y_size , &nprocs );
392
393	giet_pthread_assert( ((nprocs == 1) \|\| (nprocs == 2) \|\| (nprocs == 4)),
394	"[TRANSPOSE ERROR] number of procs per cluster must be 1, 2 or 4");
395
396	giet_pthread_assert( ((x_size == 1) \|\| (x_size == 2) \|\| (x_size == 4) \|\|
397	(x_size == 8) \|\| (x_size == 16)),
398	"[TRANSPOSE ERROR] x_size must be 1,2,4,8,16");
399
400	giet_pthread_assert( ((y_size == 1) \|\| (y_size == 2) \|\| (y_size == 4) \|\|
401	(y_size == 8) \|\| (y_size == 16)),
402	"[TRANSPOSE ERROR] y_size must be 1,2,4,8,16");
403
404	// compute number of threads
405	unsigned int nthreads = x_size * y_size * nprocs;
406
407	// shared TTY allocation
408	giet_tty_alloc( 1 );
409	lock_init( &tty_lock);
410
411	// get FBF ownership and FBF size
412	unsigned int width;
413	unsigned int height;
414	giet_fbf_alloc();
415	giet_fbf_size( &width , &height );
416
417	printf("\n[TRANSPOSE] start at cycle %d on %d cores / FBF = %d * %d pixels\n",
418	giet_proctime(), nthreads , width , height );
419
420	if ( INTERACTIVE ) // input_file_name, output_file_name, and size acquisition
421	{
422	printf("\n[TRANSPOSE] enter path for input file / default is : %s\n> ", INPUT_FILE_PATH );
423	giet_tty_gets( input_file_name , 256 );
424	printf("\n");
425	if ( strcmp( input_file_name , "" ) == 0 ) strcpy( input_file_name , INPUT_FILE_PATH );
426
427	printf("\n[TRANSPOSE] enter path for output file / default is : %s\n> ", OUTPUT_FILE_PATH );
428	giet_tty_gets( output_file_name , 256 );
429	printf("\n");
430	if ( strcmp( output_file_name , "" ) == 0 ) strcpy( output_file_name , OUTPUT_FILE_PATH );
431
432	printf("\n[TRANSPOSE] enter image size / default is : %d\n> ", IMAGE_SIZE );
433	giet_tty_getw( &image_size );
434	printf("\n");
435	if ( image_size == 0 ) image_size = IMAGE_SIZE;
436	}
437	else
438	{
439	strcpy( input_file_name , INPUT_FILE_PATH );
440	strcpy( output_file_name , OUTPUT_FILE_PATH );
441	image_size = IMAGE_SIZE;
442	}
443
444	// check image size / number of clusters
445	giet_pthread_assert( ((((image_size * image_size) / (x_size * y_size)) & 0xFFF) == 0) ,
446	"[TRANSPOSE ERROR] pixels per cluster must be multiple of 4096");
447
448	printf("\n[TRANSPOSE] input = %s / output = %s / size = %d\n",
449	input_file_name, output_file_name, image_size );
450
451	// distributed heap initialisation
452	for ( x = 0 ; x < x_size ; x++ )
453	{
454	for ( y = 0 ; y < y_size ; y++ )
455	{
456	heap_init( x , y );
457	}
458	}
459
460	// open input and output files
461	fd_in = giet_fat_open( input_file_name , O_RDONLY ); // read_only
462	if ( fd_in < 0 )
463	{
464	printf("\n[TRANSPOSE ERROR] main cannot open file %s\n", input_file_name );
465	giet_pthread_exit( NULL );
466	}
467	else
468	{
469	printf("\n[TRANSPOSE] main open file %s / fd = %d\n", input_file_name , fd_in );
470	}
471
472	fd_out = giet_fat_open( output_file_name , O_CREATE ); // create if required
473	if ( fd_out < 0 )
474	{
475	printf("\n[TRANSPOSE ERROR] main cannot open file %s\n", output_file_name );
476	giet_pthread_exit(" open() failure");
477	}
478	else
479	{
480	printf("\n[TRANSPOSE] main open file %s / fd = %d\n", output_file_name , fd_out );
481	}
482
483	// allocate thread[] array
484	pthread_t* thread = malloc( nthreads * sizeof(pthread_t) );
485
486	// barrier initialisation
487	sqt_barrier_init( &barrier, x_size , y_size , nprocs );
488
489	// Initialisation completed
490	printf("\n[TRANSPOSE] main completes initialisation\n");
491
492	// launch other threads to run execute() function
493	for ( n = 1 ; n < nthreads ; n++ )
494	{
495	if ( giet_pthread_create( &thread[n],
496	NULL, // no attribute
497	&execute,
498	NULL ) ) // no argument
499	{
500	printf("\n[TRANSPOSE ERROR] creating thread %x\n", thread[n] );
501	giet_pthread_exit( NULL );
502	}
503	}
504
505	// run the execute() function
506	execute();
507
508	// wait other threads completion
509	for ( n = 1 ; n < nthreads ; n++ )
510	{
511	if ( giet_pthread_join( thread[n], NULL ) )
512	{
513	printf("\n[TRANSPOSE ERROR] joining thread %x\n", thread[n] );
514	giet_pthread_exit( NULL );
515	}
516	else
517	{
518	printf("\n[TRANSPOSE] thread %x joined at cycle %d\n",
519	thread[n] , giet_proctime() );
520	}
521	}
522
523	// call the instrument() function
524	instrument( x_size , y_size , nprocs );
525
526	// close input and output files
527	giet_fat_close( fd_in );
528	giet_fat_close( fd_out );
529
530	// suicide
531	giet_pthread_exit( "completed" );
532
533	} // end main()
534

Note: See TracBrowser for help on using the repository browser.

Download in other formats: