Context Navigation

← Previous Revision
Next Revision →
Blame
Revision Log

transpose.c

Last change on this file was 804, checked in by alain, 8 years ago
cosmetic.
File size: 21.6 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 "string.h"
42	#include "user_barrier.h"
43	#include "malloc.h"
44
45	#define BLOCK_SIZE 512 // block size on disk
46	#define X_MAX 16 // max number of clusters in row
47	#define Y_MAX 16 // max number of clusters in column
48	#define PROCS_MAX 4 // max number of procs per cluster
49	#define CLUSTER_MAX (X_MAX * Y_MAX) // max number of clusters
50	#define IMAGE_SIZE 128 // default image size
51	#define INPUT_FILE_PATH "/misc/images_128.raw" // default input file pathname
52	#define OUTPUT_FILE_PATH "/home/trsp_128.raw" // default output file pathname
53	#define INTERACTIVE 0 // interactive capture of filenames
54	#define VERBOSE 0 // print comments on TTY
55
56
57	// macro to use a shared TTY
58	#define printf(...); { lock_acquire( &tty_lock ); \
59	giet_tty_printf(__VA_ARGS__); \
60	lock_release( &tty_lock ); }
61
62	///////////////////////////////////////////////////////
63	// global variables stored in seg_data in cluster(0,0)
64	///////////////////////////////////////////////////////
65
66	// instrumentation counters for each processor in each cluster
67	unsigned int MMAP_START[X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
68	unsigned int MMAP_END [X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
69	unsigned int TRSP_START[X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
70	unsigned int TRSP_END [X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
71	unsigned int DISP_START[X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
72	unsigned int DISP_END [X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
73
74	// arrays of pointers on distributed buffers
75	// one input buffer & one output buffer per cluster
76	unsigned char* buf_in [CLUSTER_MAX];
77	unsigned char* buf_out[CLUSTER_MAX];
78
79	// lock protecting shared TTY
80	user_lock_t tty_lock;
81
82	// synchronisation barrier (all threads)
83	giet_sqt_barrier_t barrier;
84
85	// input & output files pathname and size
86	char input_file_name[256];
87	char output_file_name[256];
88	unsigned int image_size;
89
90	// input & output file descriptors
91	int fd_in;
92	int fd_out;
93
94	////////////////////////////////////////////////////////////////
95	__attribute__ ((constructor)) void execute( unsigned int* tid )
96	////////////////////////////////////////////////////////////////
97	{
98	unsigned int l; // line index for loops
99	unsigned int p; // pixel index for loops
100
101	// get processor identifiers
102	unsigned int x_id; // x cluster coordinate
103	unsigned int y_id; // y cluster coordinate
104	unsigned int p_id; // local processor index
105
106	giet_proc_xyp( &x_id, &y_id, &p_id);
107
108	// get & check plat-form parameters
109	unsigned int x_size; // number of clusters in a row
110	unsigned int y_size; // number of clusters in a column
111	unsigned int nprocs; // number of processors per cluster
112
113	giet_procs_number( &x_size , &y_size , &nprocs );
114
115	unsigned int nclusters = x_size * y_size; // number of clusters
116	unsigned int nthreads = x_size * y_size * nprocs; // number of threads
117	unsigned int npixels = image_size * image_size; // pixels per image
118	unsigned int cluster_id = (x_id * y_size) + y_id; // "continuous" index
119	unsigned int thread_id = (cluster_id * nprocs) + p_id; // "continuous" index
120
121	if ( VERBOSE )
122	{
123	printf("\n@@@ thread %d / %d running on P[%d,%d,%d]\n",
124	*tid , nthreads , x_id , y_id , p_id );
125	}
126
127	// parallel load of image:
128	// thread running on processor[x,y,0]
129	// map input & output files in buf_in & buf_out buffers.
130
131	MMAP_START[x_id][y_id][p_id] = giet_proctime();
132
133	if ( p_id == 0 )
134	{
135	// map buf_in
136	unsigned int length = npixels / nclusters;
137	unsigned int offset = length * cluster_id;
138
139	buf_in[cluster_id] = giet_fat_mmap( NULL,
140	length,
141	MAP_PROT_READ,
142	MAP_SHARED,
143	fd_in,
144	offset );
145	if ( buf_in[cluster_id] == NULL )
146	{
147	printf("\n[TRANSPOSE ERROR] Thread[%d,%d,%d] cannot map input file\n",
148	x_id , y_id , p_id );
149	giet_pthread_exit( NULL );
150	}
151
152	if ( VERBOSE )
153	{
154	printf("\n@@@ thread %d running on P[%d,%d,%d] call mmap for input file\n"
155	" length = %x / offset = %x / buf_in = %x\n",
156	*tid , x_id , y_id , p_id , length , offset , buf_in[cluster_id] );
157	}
158
159	// map buf_out
160	buf_out[cluster_id] = giet_fat_mmap( NULL,
161	length,
162	MAP_PROT_WRITE,
163	MAP_SHARED,
164	fd_out,
165	offset );
166
167	if ( buf_out[cluster_id] == NULL )
168	{
169	printf("\n[TRANSPOSE ERROR] Thread[%d,%d,%d] cannot map output file\n",
170	x_id , y_id , p_id );
171	giet_pthread_exit( NULL );
172	}
173
174	if ( VERBOSE )
175	{
176	printf("\n@@@ thread %d running on P[%d,%d,%d] call mmap for output file\n"
177	" length = %x / offset = %x / buf_out = %x\n",
178	*tid , x_id , y_id , p_id , length , offset , buf_out[cluster_id] );
179	}
180
181	}
182
183	MMAP_END[x_id][y_id][p_id] = giet_proctime();
184
185	/////////////////////////////
186	sqt_barrier_wait( &barrier );
187	/////////////////////////////
188
189	// parallel transpose from buf_in to buf_out
190	// each thread makes the transposition for nlt lines (nlt = image_size/nthreads)
191	// from line [thread_idnlt] to line [(thread_id + 1)nlt - 1]
192	// (p,l) are the absolute pixel coordinates in the source image
193
194	TRSP_START[x_id][y_id][p_id] = giet_proctime();
195
196	unsigned int nlt = image_size / nthreads; // number of lines per thread
197	unsigned int nlc = image_size / nclusters; // number of lines per cluster
198
199	unsigned int src_cluster;
200	unsigned int src_index;
201	unsigned int dst_cluster;
202	unsigned int dst_index;
203
204	unsigned char byte;
205
206	unsigned int first = thread_id * nlt; // first line index for a given thread
207	unsigned int last = first + nlt; // last line index for a given thread
208
209	for ( l = first ; l < last ; l++ )
210	{
211	// in each iteration we transfer one byte
212	for ( p = 0 ; p < image_size ; p++ )
213	{
214	// read one byte from local buf_in
215	src_cluster = l / nlc;
216	src_index = (l % nlc)*image_size + p;
217	byte = buf_in[src_cluster][src_index];
218
219	// write one byte to remote buf_out
220	dst_cluster = p / nlc;
221	dst_index = (p % nlc)*image_size + l;
222	buf_out[dst_cluster][dst_index] = byte;
223	}
224	}
225
226	if ( (p_id == 0) && (x_id==0) && (y_id==0) )
227	{
228	printf("\n[TRANSPOSE] Thread[%d,%d,%d] completes transpose at cycle %d\n",
229	x_id, y_id, p_id, giet_proctime() );
230	}
231
232	TRSP_END[x_id][y_id][p_id] = giet_proctime();
233
234	/////////////////////////////
235	sqt_barrier_wait( &barrier );
236	/////////////////////////////
237
238	// parallel display from local buf_out to frame buffer
239	// all threads contribute to display using memcpy...
240
241	DISP_START[x_id][y_id][p_id] = giet_proctime();
242
243	unsigned int npt = npixels / nthreads; // number of pixels per thread
244
245	giet_fbf_sync_write( npt * thread_id,
246	&buf_out[cluster_id][p_id*npt],
247	npt );
248
249	if ( (x_id==0) && (y_id==0) && (p_id==0) )
250	{
251	printf("\n[TRANSPOSE] Thread[%d,%d,%d] completes display at cycle %d\n",
252	x_id, y_id, p_id, giet_proctime() );
253	}
254
255	DISP_END[x_id][y_id][p_id] = giet_proctime();
256
257	/////////////////////////////
258	sqt_barrier_wait( &barrier );
259	/////////////////////////////
260
261	// all threads, but thread[0,0,0], suicide
262	if ( (x_id != 0) \|\| (y_id != 0) \|\| (p_id != 0) )
263	giet_pthread_exit( "completed" );
264
265	} // end execute()
266
267
268
269	//////////////////////////////////////
270	void instrument( unsigned int x_size,
271	unsigned int y_size,
272	unsigned int nprocs )
273	//////////////////////////////////////
274	{
275	unsigned int x, y, l;
276
277	unsigned int min_load_start = 0xFFFFFFFF;
278	unsigned int max_load_start = 0;
279	unsigned int min_load_ended = 0xFFFFFFFF;
280	unsigned int max_load_ended = 0;
281	unsigned int min_trsp_start = 0xFFFFFFFF;
282	unsigned int max_trsp_start = 0;
283	unsigned int min_trsp_ended = 0xFFFFFFFF;
284	unsigned int max_trsp_ended = 0;
285	unsigned int min_disp_start = 0xFFFFFFFF;
286	unsigned int max_disp_start = 0;
287	unsigned int min_disp_ended = 0xFFFFFFFF;
288	unsigned int max_disp_ended = 0;
289
290	char string[64];
291
292	snprintf( string , 64 , "/home/transpose_%d_%d_%d" , x_size , y_size , nprocs );
293
294	// open instrumentation file
295	unsigned int fd = giet_fat_open( string , O_CREAT );
296	if ( fd < 0 )
297	{
298	printf("\n[TRANSPOSE ERROR] cannot open instrumentation file %s\n", string );
299	giet_pthread_exit( NULL );
300	}
301
302	for (x = 0; x < x_size; x++)
303	{
304	for (y = 0; y < y_size; y++)
305	{
306	for ( l = 0 ; l < nprocs ; l++ )
307	{
308	if (MMAP_START[x][y][l] < min_load_start) min_load_start = MMAP_START[x][y][l];
309	if (MMAP_START[x][y][l] > max_load_start) max_load_start = MMAP_START[x][y][l];
310	if (MMAP_END[x][y][l] < min_load_ended) min_load_ended = MMAP_END[x][y][l];
311	if (MMAP_END[x][y][l] > max_load_ended) max_load_ended = MMAP_END[x][y][l];
312	if (TRSP_START[x][y][l] < min_trsp_start) min_trsp_start = TRSP_START[x][y][l];
313	if (TRSP_START[x][y][l] > max_trsp_start) max_trsp_start = TRSP_START[x][y][l];
314	if (TRSP_END[x][y][l] < min_trsp_ended) min_trsp_ended = TRSP_END[x][y][l];
315	if (TRSP_END[x][y][l] > max_trsp_ended) max_trsp_ended = TRSP_END[x][y][l];
316	if (DISP_START[x][y][l] < min_disp_start) min_disp_start = DISP_START[x][y][l];
317	if (DISP_START[x][y][l] > max_disp_start) max_disp_start = DISP_START[x][y][l];
318	if (DISP_END[x][y][l] < min_disp_ended) min_disp_ended = DISP_END[x][y][l];
319	if (DISP_END[x][y][l] > max_disp_ended) max_disp_ended = DISP_END[x][y][l];
320	}
321	}
322	}
323
324	giet_tty_printf( "\n ------ %s ------\n" , string );
325	giet_fat_fprintf( fd , "\n ------ %s ------\n" , string );
326
327	giet_tty_printf( " - MMAP_START : min = %d / max = %d / med = %d / delta = %d\n",
328	min_load_start, max_load_start, (min_load_start+max_load_start)/2,
329	max_load_start-min_load_start );
330
331	giet_fat_fprintf( fd , " - MMAP_START : min = %d / max = %d / med = %d / delta = %d\n",
332	min_load_start, max_load_start, (min_load_start+max_load_start)/2,
333	max_load_start-min_load_start );
334
335	giet_tty_printf( " - MMAP_END : min = %d / max = %d / med = %d / delta = %d\n",
336	min_load_ended, max_load_ended, (min_load_ended+max_load_ended)/2,
337	max_load_ended-min_load_ended );
338
339	giet_fat_fprintf( fd , " - MMAP_END : min = %d / max = %d / med = %d / delta = %d\n",
340	min_load_ended, max_load_ended, (min_load_ended+max_load_ended)/2,
341	max_load_ended-min_load_ended );
342
343	giet_tty_printf( " - TRSP_START : min = %d / max = %d / med = %d / delta = %d\n",
344	min_trsp_start, max_trsp_start, (min_trsp_start+max_trsp_start)/2,
345	max_trsp_start-min_trsp_start );
346
347	giet_fat_fprintf( fd , " - TRSP_START : min = %d / max = %d / med = %d / delta = %d\n",
348	min_trsp_start, max_trsp_start, (min_trsp_start+max_trsp_start)/2,
349	max_trsp_start-min_trsp_start );
350
351	giet_tty_printf( " - TRSP_END : min = %d / max = %d / med = %d / delta = %d\n",
352	min_trsp_ended, max_trsp_ended, (min_trsp_ended+max_trsp_ended)/2,
353	max_trsp_ended-min_trsp_ended );
354
355	giet_fat_fprintf( fd , " - TRSP_END : min = %d / max = %d / med = %d / delta = %d\n",
356	min_trsp_ended, max_trsp_ended, (min_trsp_ended+max_trsp_ended)/2,
357	max_trsp_ended-min_trsp_ended );
358
359	giet_tty_printf( " - DISP_START : min = %d / max = %d / med = %d / delta = %d\n",
360	min_disp_start, max_disp_start, (min_disp_start+max_disp_start)/2,
361	max_disp_start-min_disp_start );
362
363	giet_fat_fprintf( fd , " - DISP_START : min = %d / max = %d / med = %d / delta = %d\n",
364	min_disp_start, max_disp_start, (min_disp_start+max_disp_start)/2,
365	max_disp_start-min_disp_start );
366
367	giet_tty_printf( " - DISP_END : min = %d / max = %d / med = %d / delta = %d\n",
368	min_disp_ended, max_disp_ended, (min_disp_ended+max_disp_ended)/2,
369	max_disp_ended-min_disp_ended );
370
371	giet_fat_fprintf( fd , " - DISP_END : min = %d / max = %d / med = %d / delta = %d\n",
372	min_disp_ended, max_disp_ended, (min_disp_ended+max_disp_ended)/2,
373	max_disp_ended-min_disp_ended );
374
375	giet_fat_close( fd );
376
377	} // end instrument()
378
379
380
381	//////////////////////////////////////////
382	__attribute__ ((constructor)) void main()
383	//////////////////////////////////////////
384	{
385	// indexes for loops
386	unsigned int x , y , n;
387
388	// get identifiers for proc executing main
389	unsigned int x_id; // x cluster coordinate
390	unsigned int y_id; // y cluster coordinate
391	unsigned int p_id; // local processor index
392	giet_proc_xyp( &x_id , &y_id , &p_id );
393
394	// get & check plat-form parameters
395	unsigned int x_size; // number of clusters in a row
396	unsigned int y_size; // number of clusters in a column
397	unsigned int nprocs; // number of processors per cluster
398	giet_procs_number( &x_size , &y_size , &nprocs );
399
400	giet_pthread_assert( ((nprocs == 1) \|\| (nprocs == 2) \|\| (nprocs == 4)),
401	"[TRANSPOSE ERROR] number of procs per cluster must be 1, 2 or 4");
402
403	giet_pthread_assert( ((x_size == 1) \|\| (x_size == 2) \|\| (x_size == 4) \|\|
404	(x_size == 8) \|\| (x_size == 16)),
405	"[TRANSPOSE ERROR] x_size must be 1,2,4,8,16");
406
407	giet_pthread_assert( ((y_size == 1) \|\| (y_size == 2) \|\| (y_size == 4) \|\|
408	(y_size == 8) \|\| (y_size == 16)),
409	"[TRANSPOSE ERROR] y_size must be 1,2,4,8,16");
410
411	// compute number of threads
412	unsigned int nthreads = x_size * y_size * nprocs;
413
414	// shared TTY allocation
415	giet_tty_alloc( 1 );
416	lock_init( &tty_lock);
417
418	// get FBF ownership and FBF size
419	unsigned int width;
420	unsigned int height;
421	giet_fbf_alloc();
422	giet_fbf_size( &width , &height );
423
424	printf("\n[TRANSPOSE] start at cycle %d on %d cores / FBF = %d * %d pixels\n",
425	giet_proctime(), nthreads , width , height );
426
427	if ( INTERACTIVE ) // input_file_name, output_file_name, and size acquisition
428	{
429	printf("\n[TRANSPOSE] enter path for input file / default is : %s\n> ", INPUT_FILE_PATH );
430	giet_tty_gets( input_file_name , 256 );
431	printf("\n");
432	if ( strcmp( input_file_name , "" ) == 0 ) strcpy( input_file_name , INPUT_FILE_PATH );
433
434	printf("\n[TRANSPOSE] enter path for output file / default is : %s\n> ", OUTPUT_FILE_PATH );
435	giet_tty_gets( output_file_name , 256 );
436	printf("\n");
437	if ( strcmp( output_file_name , "" ) == 0 ) strcpy( output_file_name , OUTPUT_FILE_PATH );
438
439	printf("\n[TRANSPOSE] enter image size / default is : %d\n> ", IMAGE_SIZE );
440	giet_tty_getw( &image_size );
441	printf("\n");
442	if ( image_size == 0 ) image_size = IMAGE_SIZE;
443	}
444	else
445	{
446	strcpy( input_file_name , INPUT_FILE_PATH );
447	strcpy( output_file_name , OUTPUT_FILE_PATH );
448	image_size = IMAGE_SIZE;
449	}
450
451	// check image size / number of clusters
452	giet_pthread_assert( ((((image_size * image_size) / (x_size * y_size)) & 0xFFF) == 0) ,
453	"[TRANSPOSE ERROR] pixels per cluster must be multiple of 4096");
454
455	printf("\n[TRANSPOSE] input = %s / output = %s / size = %d\n",
456	input_file_name, output_file_name, image_size );
457
458	// distributed heap initialisation
459	for ( x = 0 ; x < x_size ; x++ )
460	{
461	for ( y = 0 ; y < y_size ; y++ )
462	{
463	heap_init( x , y );
464	}
465	}
466
467	// open input and output files
468	fd_in = giet_fat_open( input_file_name , O_RDONLY ); // read_only
469	if ( fd_in < 0 )
470	{
471	printf("\n[TRANSPOSE ERROR] main cannot open file %s\n", input_file_name );
472	giet_pthread_exit( NULL );
473	}
474	else
475	{
476	printf("\n[TRANSPOSE] main open file %s / fd = %d\n", input_file_name , fd_in );
477	}
478
479	fd_out = giet_fat_open( output_file_name , O_CREAT ); // create if required
480	if ( fd_out < 0 )
481	{
482	printf("\n[TRANSPOSE ERROR] main cannot open file %s\n", output_file_name );
483	giet_pthread_exit(" open() failure");
484	}
485	else
486	{
487	printf("\n[TRANSPOSE] main open file %s / fd = %d\n", output_file_name , fd_out );
488	}
489
490	// allocate thread[] array : thread identidiers defined by the kernel
491	pthread_t* thread = malloc( nthreads * sizeof(pthread_t) );
492
493	// allocate tid[] array : continuous thred index defined by the user
494	unsigned int* tid = malloc( nthreads * sizeof(unsigned int) );
495
496	// barrier initialisation
497	sqt_barrier_init( &barrier, x_size , y_size , nprocs );
498
499	// Initialisation completed
500	printf("\n[TRANSPOSE] main completes initialisation\n");
501
502	// launch other threads to run execute() function
503	for ( n = 1 ; n < nthreads ; n++ )
504	{
505	tid[n] = n;
506	if ( giet_pthread_create( &thread[n],
507	NULL, // no attribute
508	&execute,
509	&tid[n] ) ) // pointer on thread index
510	{
511	printf("\n[TRANSPOSE ERROR] creating thread %d / id = %x\n", tid[n] , thread[n] );
512	giet_pthread_exit( NULL );
513	}
514	}
515
516	// run the execute() function
517	execute( &tid[0] );
518
519	// wait other threads completion
520	for ( n = 1 ; n < nthreads ; n++ )
521	{
522	if ( giet_pthread_join( thread[n], NULL ) )
523	{
524	printf("\n[TRANSPOSE ERROR] joining thread %d / id = %x\n", tid[n] , thread[n] );
525	giet_pthread_exit( NULL );
526	}
527	else
528	{
529	printf("\n[TRANSPOSE] thread %x joined at cycle %d\n",
530	thread[n] , giet_proctime() );
531	}
532	}
533
534	// call the instrument() function
535	instrument( x_size , y_size , nprocs );
536
537	// close input and output files
538	giet_fat_close( fd_in );
539	giet_fat_close( fd_out );
540
541	// suicide
542	giet_pthread_exit( "completed" );
543
544	} // end main()
545

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

Context Navigation

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

Download in other formats: