Context Navigation

source: soft/giet_vm/applications/transpose/transpose.c @ 801

Last change on this file since 801 was 782, checked in by alain, 9 years ago
1) Introduce the string library. 2) Introduce a heap in the shell application.
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 "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()
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	// parallel load of image:
122	// thread running on processor[x,y,0]
123	// map input & output files in buf_in & buf_out buffers.
124
125	MMAP_START[x_id][y_id][p_id] = giet_proctime();
126
127	if ( p_id == 0 )
128	{
129	// map buf_in
130	unsigned int length = npixels / nclusters;
131	unsigned int offset = length * cluster_id;
132
133	buf_in[cluster_id] = giet_fat_mmap( NULL,
134	length,
135	MAP_PROT_READ,
136	MAP_SHARED,
137	fd_in,
138	offset );
139	if ( buf_in[cluster_id] == NULL )
140	{
141	printf("\n[TRANSPOSE ERROR] Thread[%d,%d,%d] cannot map input file\n",
142	x_id , y_id , p_id );
143	giet_pthread_exit( NULL );
144	}
145
146	if ( VERBOSE )
147	{
148	printf("\n@@@ Thread[%d,%d,%d] call mmap for input file\n"
149	" length = %x / offset = %x / buf_in = %x\n",
150	x_id , y_id , p_id , length , offset , buf_in[cluster_id] );
151	}
152
153	// map buf_out
154	buf_out[cluster_id] = giet_fat_mmap( NULL,
155	length,
156	MAP_PROT_WRITE,
157	MAP_SHARED,
158	fd_out,
159	offset );
160
161	if ( buf_out[cluster_id] == NULL )
162	{
163	printf("\n[TRANSPOSE ERROR] Thread[%d,%d,%d] cannot map output file\n",
164	x_id , y_id , p_id );
165	giet_pthread_exit( NULL );
166	}
167
168	if ( VERBOSE )
169	{
170	printf("\n@@@ Thread[%d,%d,%d] call mmap for output file\n"
171	" length = %x / offset = %x / buf_out = %x\n",
172	x_id , y_id , p_id , length , offset , buf_out[cluster_id] );
173	}
174
175	}
176
177	MMAP_END[x_id][y_id][p_id] = giet_proctime();
178
179	/////////////////////////////
180	sqt_barrier_wait( &barrier );
181	/////////////////////////////
182
183	// parallel transpose from buf_in to buf_out
184	// each thread makes the transposition for nlt lines (nlt = image_size/nthreads)
185	// from line [thread_idnlt] to line [(thread_id + 1)nlt - 1]
186	// (p,l) are the absolute pixel coordinates in the source image
187
188	TRSP_START[x_id][y_id][p_id] = giet_proctime();
189
190	unsigned int nlt = image_size / nthreads; // number of lines per thread
191	unsigned int nlc = image_size / nclusters; // number of lines per cluster
192
193	unsigned int src_cluster;
194	unsigned int src_index;
195	unsigned int dst_cluster;
196	unsigned int dst_index;
197
198	unsigned char byte;
199
200	unsigned int first = thread_id * nlt; // first line index for a given thread
201	unsigned int last = first + nlt; // last line index for a given thread
202
203	for ( l = first ; l < last ; l++ )
204	{
205	// in each iteration we transfer one byte
206	for ( p = 0 ; p < image_size ; p++ )
207	{
208	// read one byte from local buf_in
209	src_cluster = l / nlc;
210	src_index = (l % nlc)*image_size + p;
211	byte = buf_in[src_cluster][src_index];
212
213	// write one byte to remote buf_out
214	dst_cluster = p / nlc;
215	dst_index = (p % nlc)*image_size + l;
216	buf_out[dst_cluster][dst_index] = byte;
217	}
218	}
219
220	if ( (p_id == 0) && (x_id==0) && (y_id==0) )
221	{
222	printf("\n[TRANSPOSE] Thread[%d,%d,%d] completes transpose at cycle %d\n",
223	x_id, y_id, p_id, giet_proctime() );
224	}
225
226	TRSP_END[x_id][y_id][p_id] = giet_proctime();
227
228	/////////////////////////////
229	sqt_barrier_wait( &barrier );
230	/////////////////////////////
231
232	// parallel display from local buf_out to frame buffer
233	// all threads contribute to display using memcpy...
234
235	DISP_START[x_id][y_id][p_id] = giet_proctime();
236
237	unsigned int npt = npixels / nthreads; // number of pixels per thread
238
239	giet_fbf_sync_write( npt * thread_id,
240	&buf_out[cluster_id][p_id*npt],
241	npt );
242
243	if ( (x_id==0) && (y_id==0) && (p_id==0) )
244	{
245	printf("\n[TRANSPOSE] Thread[%d,%d,%d] completes display at cycle %d\n",
246	x_id, y_id, p_id, giet_proctime() );
247	}
248
249	DISP_END[x_id][y_id][p_id] = giet_proctime();
250
251	/////////////////////////////
252	sqt_barrier_wait( &barrier );
253	/////////////////////////////
254
255	// all threads, but thread[0,0,0], suicide
256	if ( (x_id != 0) \|\| (y_id != 0) \|\| (p_id != 0) )
257	giet_pthread_exit( "completed" );
258
259	} // end execute()
260
261
262
263	//////////////////////////////////////
264	void instrument( unsigned int x_size,
265	unsigned int y_size,
266	unsigned int nprocs )
267	//////////////////////////////////////
268	{
269	unsigned int x, y, l;
270
271	unsigned int min_load_start = 0xFFFFFFFF;
272	unsigned int max_load_start = 0;
273	unsigned int min_load_ended = 0xFFFFFFFF;
274	unsigned int max_load_ended = 0;
275	unsigned int min_trsp_start = 0xFFFFFFFF;
276	unsigned int max_trsp_start = 0;
277	unsigned int min_trsp_ended = 0xFFFFFFFF;
278	unsigned int max_trsp_ended = 0;
279	unsigned int min_disp_start = 0xFFFFFFFF;
280	unsigned int max_disp_start = 0;
281	unsigned int min_disp_ended = 0xFFFFFFFF;
282	unsigned int max_disp_ended = 0;
283
284	char string[64];
285
286	snprintf( string , 64 , "/home/transpose_%d_%d_%d" , x_size , y_size , nprocs );
287
288	// open instrumentation file
289	unsigned int fd = giet_fat_open( string , O_CREAT );
290	if ( fd < 0 )
291	{
292	printf("\n[TRANSPOSE ERROR] cannot open instrumentation file %s\n", string );
293	giet_pthread_exit( NULL );
294	}
295
296	for (x = 0; x < x_size; x++)
297	{
298	for (y = 0; y < y_size; y++)
299	{
300	for ( l = 0 ; l < nprocs ; l++ )
301	{
302	if (MMAP_START[x][y][l] < min_load_start) min_load_start = MMAP_START[x][y][l];
303	if (MMAP_START[x][y][l] > max_load_start) max_load_start = MMAP_START[x][y][l];
304	if (MMAP_END[x][y][l] < min_load_ended) min_load_ended = MMAP_END[x][y][l];
305	if (MMAP_END[x][y][l] > max_load_ended) max_load_ended = MMAP_END[x][y][l];
306	if (TRSP_START[x][y][l] < min_trsp_start) min_trsp_start = TRSP_START[x][y][l];
307	if (TRSP_START[x][y][l] > max_trsp_start) max_trsp_start = TRSP_START[x][y][l];
308	if (TRSP_END[x][y][l] < min_trsp_ended) min_trsp_ended = TRSP_END[x][y][l];
309	if (TRSP_END[x][y][l] > max_trsp_ended) max_trsp_ended = TRSP_END[x][y][l];
310	if (DISP_START[x][y][l] < min_disp_start) min_disp_start = DISP_START[x][y][l];
311	if (DISP_START[x][y][l] > max_disp_start) max_disp_start = DISP_START[x][y][l];
312	if (DISP_END[x][y][l] < min_disp_ended) min_disp_ended = DISP_END[x][y][l];
313	if (DISP_END[x][y][l] > max_disp_ended) max_disp_ended = DISP_END[x][y][l];
314	}
315	}
316	}
317
318	giet_tty_printf( "\n ------ %s ------\n" , string );
319	giet_fat_fprintf( fd , "\n ------ %s ------\n" , string );
320
321	giet_tty_printf( " - MMAP_START : min = %d / max = %d / med = %d / delta = %d\n",
322	min_load_start, max_load_start, (min_load_start+max_load_start)/2,
323	max_load_start-min_load_start );
324
325	giet_fat_fprintf( fd , " - MMAP_START : min = %d / max = %d / med = %d / delta = %d\n",
326	min_load_start, max_load_start, (min_load_start+max_load_start)/2,
327	max_load_start-min_load_start );
328
329	giet_tty_printf( " - MMAP_END : min = %d / max = %d / med = %d / delta = %d\n",
330	min_load_ended, max_load_ended, (min_load_ended+max_load_ended)/2,
331	max_load_ended-min_load_ended );
332
333	giet_fat_fprintf( fd , " - MMAP_END : min = %d / max = %d / med = %d / delta = %d\n",
334	min_load_ended, max_load_ended, (min_load_ended+max_load_ended)/2,
335	max_load_ended-min_load_ended );
336
337	giet_tty_printf( " - TRSP_START : min = %d / max = %d / med = %d / delta = %d\n",
338	min_trsp_start, max_trsp_start, (min_trsp_start+max_trsp_start)/2,
339	max_trsp_start-min_trsp_start );
340
341	giet_fat_fprintf( fd , " - TRSP_START : min = %d / max = %d / med = %d / delta = %d\n",
342	min_trsp_start, max_trsp_start, (min_trsp_start+max_trsp_start)/2,
343	max_trsp_start-min_trsp_start );
344
345	giet_tty_printf( " - TRSP_END : min = %d / max = %d / med = %d / delta = %d\n",
346	min_trsp_ended, max_trsp_ended, (min_trsp_ended+max_trsp_ended)/2,
347	max_trsp_ended-min_trsp_ended );
348
349	giet_fat_fprintf( fd , " - TRSP_END : min = %d / max = %d / med = %d / delta = %d\n",
350	min_trsp_ended, max_trsp_ended, (min_trsp_ended+max_trsp_ended)/2,
351	max_trsp_ended-min_trsp_ended );
352
353	giet_tty_printf( " - DISP_START : min = %d / max = %d / med = %d / delta = %d\n",
354	min_disp_start, max_disp_start, (min_disp_start+max_disp_start)/2,
355	max_disp_start-min_disp_start );
356
357	giet_fat_fprintf( fd , " - DISP_START : min = %d / max = %d / med = %d / delta = %d\n",
358	min_disp_start, max_disp_start, (min_disp_start+max_disp_start)/2,
359	max_disp_start-min_disp_start );
360
361	giet_tty_printf( " - DISP_END : min = %d / max = %d / med = %d / delta = %d\n",
362	min_disp_ended, max_disp_ended, (min_disp_ended+max_disp_ended)/2,
363	max_disp_ended-min_disp_ended );
364
365	giet_fat_fprintf( fd , " - DISP_END : min = %d / max = %d / med = %d / delta = %d\n",
366	min_disp_ended, max_disp_ended, (min_disp_ended+max_disp_ended)/2,
367	max_disp_ended-min_disp_ended );
368
369	giet_fat_close( fd );
370
371	} // end instrument()
372
373
374
375	//////////////////////////////////////////
376	__attribute__ ((constructor)) void main()
377	//////////////////////////////////////////
378	{
379	// indexes for loops
380	unsigned int x , y , n;
381
382	// get identifiers for proc executing main
383	unsigned int x_id; // x cluster coordinate
384	unsigned int y_id; // y cluster coordinate
385	unsigned int p_id; // local processor index
386	giet_proc_xyp( &x_id , &y_id , &p_id );
387
388	// get & check plat-form parameters
389	unsigned int x_size; // number of clusters in a row
390	unsigned int y_size; // number of clusters in a column
391	unsigned int nprocs; // number of processors per cluster
392	giet_procs_number( &x_size , &y_size , &nprocs );
393
394	giet_pthread_assert( ((nprocs == 1) \|\| (nprocs == 2) \|\| (nprocs == 4)),
395	"[TRANSPOSE ERROR] number of procs per cluster must be 1, 2 or 4");
396
397	giet_pthread_assert( ((x_size == 1) \|\| (x_size == 2) \|\| (x_size == 4) \|\|
398	(x_size == 8) \|\| (x_size == 16)),
399	"[TRANSPOSE ERROR] x_size must be 1,2,4,8,16");
400
401	giet_pthread_assert( ((y_size == 1) \|\| (y_size == 2) \|\| (y_size == 4) \|\|
402	(y_size == 8) \|\| (y_size == 16)),
403	"[TRANSPOSE ERROR] y_size must be 1,2,4,8,16");
404
405	// compute number of threads
406	unsigned int nthreads = x_size * y_size * nprocs;
407
408	// shared TTY allocation
409	giet_tty_alloc( 1 );
410	lock_init( &tty_lock);
411
412	// get FBF ownership and FBF size
413	unsigned int width;
414	unsigned int height;
415	giet_fbf_alloc();
416	giet_fbf_size( &width , &height );
417
418	printf("\n[TRANSPOSE] start at cycle %d on %d cores / FBF = %d * %d pixels\n",
419	giet_proctime(), nthreads , width , height );
420
421	if ( INTERACTIVE ) // input_file_name, output_file_name, and size acquisition
422	{
423	printf("\n[TRANSPOSE] enter path for input file / default is : %s\n> ", INPUT_FILE_PATH );
424	giet_tty_gets( input_file_name , 256 );
425	printf("\n");
426	if ( strcmp( input_file_name , "" ) == 0 ) strcpy( input_file_name , INPUT_FILE_PATH );
427
428	printf("\n[TRANSPOSE] enter path for output file / default is : %s\n> ", OUTPUT_FILE_PATH );
429	giet_tty_gets( output_file_name , 256 );
430	printf("\n");
431	if ( strcmp( output_file_name , "" ) == 0 ) strcpy( output_file_name , OUTPUT_FILE_PATH );
432
433	printf("\n[TRANSPOSE] enter image size / default is : %d\n> ", IMAGE_SIZE );
434	giet_tty_getw( &image_size );
435	printf("\n");
436	if ( image_size == 0 ) image_size = IMAGE_SIZE;
437	}
438	else
439	{
440	strcpy( input_file_name , INPUT_FILE_PATH );
441	strcpy( output_file_name , OUTPUT_FILE_PATH );
442	image_size = IMAGE_SIZE;
443	}
444
445	// check image size / number of clusters
446	giet_pthread_assert( ((((image_size * image_size) / (x_size * y_size)) & 0xFFF) == 0) ,
447	"[TRANSPOSE ERROR] pixels per cluster must be multiple of 4096");
448
449	printf("\n[TRANSPOSE] input = %s / output = %s / size = %d\n",
450	input_file_name, output_file_name, image_size );
451
452	// distributed heap initialisation
453	for ( x = 0 ; x < x_size ; x++ )
454	{
455	for ( y = 0 ; y < y_size ; y++ )
456	{
457	heap_init( x , y );
458	}
459	}
460
461	// open input and output files
462	fd_in = giet_fat_open( input_file_name , O_RDONLY ); // read_only
463	if ( fd_in < 0 )
464	{
465	printf("\n[TRANSPOSE ERROR] main cannot open file %s\n", input_file_name );
466	giet_pthread_exit( NULL );
467	}
468	else
469	{
470	printf("\n[TRANSPOSE] main open file %s / fd = %d\n", input_file_name , fd_in );
471	}
472
473	fd_out = giet_fat_open( output_file_name , O_CREAT ); // create if required
474	if ( fd_out < 0 )
475	{
476	printf("\n[TRANSPOSE ERROR] main cannot open file %s\n", output_file_name );
477	giet_pthread_exit(" open() failure");
478	}
479	else
480	{
481	printf("\n[TRANSPOSE] main open file %s / fd = %d\n", output_file_name , fd_out );
482	}
483
484	// allocate thread[] array
485	pthread_t* thread = malloc( nthreads * sizeof(pthread_t) );
486
487	// barrier initialisation
488	sqt_barrier_init( &barrier, x_size , y_size , nprocs );
489
490	// Initialisation completed
491	printf("\n[TRANSPOSE] main completes initialisation\n");
492
493	// launch other threads to run execute() function
494	for ( n = 1 ; n < nthreads ; n++ )
495	{
496	if ( giet_pthread_create( &thread[n],
497	NULL, // no attribute
498	&execute,
499	NULL ) ) // no argument
500	{
501	printf("\n[TRANSPOSE ERROR] creating thread %x\n", thread[n] );
502	giet_pthread_exit( NULL );
503	}
504	}
505
506	// run the execute() function
507	execute();
508
509	// wait other threads completion
510	for ( n = 1 ; n < nthreads ; n++ )
511	{
512	if ( giet_pthread_join( thread[n], NULL ) )
513	{
514	printf("\n[TRANSPOSE ERROR] joining thread %x\n", thread[n] );
515	giet_pthread_exit( NULL );
516	}
517	else
518	{
519	printf("\n[TRANSPOSE] thread %x joined at cycle %d\n",
520	thread[n] , giet_proctime() );
521	}
522	}
523
524	// call the instrument() function
525	instrument( x_size , y_size , nprocs );
526
527	// close input and output files
528	giet_fat_close( fd_in );
529	giet_fat_close( fd_out );
530
531	// suicide
532	giet_pthread_exit( "completed" );
533
534	} // end main()
535

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

Download in other formats: